The next engineering advantage is how you run agents.
Engineering output is accelerating, and AI now reaches more of the work that ships. The next step is to turn that capacity into a system teams can trust, measure and improve.
This notebook holds the evidence behind the report. Each chapter sets out the argument, the data tables behind it and the limits on what the numbers can support.
It analyzes 22,623 engineers with merged output in Q3 2026, with monthly Index data through September and a separate view of coding-agent economics.
Weave is trusted by 500+ engineering organizations, from startups to Fortune 100 and Fortune 500 companies. That gives us a direct view into the systems teams use to build software: their source code, reviews, coding agents and AI spend.
What the data says
AI assistance has spread past the early adopters. Observed adoption reached 82%, and in September the median organization’s AI-assisted share of merged output reached 53%.
Output rose 87% from June to September. A fixed cohort of organizations also shows substantial growth: the median change from Q3 2025 to Q3 2026 was 3.1×. New organizations joining the platform do not explain all of the acceleration.
The economics are less settled. From June to September, recorded spend grew about 2.5× while output grew about 1.9×. In September, matched cost per output improved 14.8%. We treat that as a signal to follow. It is too early to call it a trend.
The next operating problem is conversion. In the Q3 spend cut, $40 of every $100 was linked to merged code, and $23 of that also had a clean completion signal. The cut excludes one very large spender, so read it with that boundary attached.
Our conclusion
Model capability is only part of the engineering system. Teams can build a repeatable advantage in task design, codified workflows, repository context, acceptance checks, review and session economics.
01 AI in the workflow
AI is becoming part of ordinary engineering work.
The important change is how deeply AI reaches into the work that ships.
In Q3 we observed AI activity for 18,598 of the 22,623 engineers with merged output in the report. That is 82%, up from 76% in Q2. Adoption is still worth asking about, but the answer tells you less on its own than it used to.
The deeper change shows up in output. By September, 57% of pooled merged output value came from PRs with detected AI assistance. The median organization reached 53%, up from 37% in June. The gap between the pooled result and the median organization narrowed from 15 percentage points to four.
That narrowing matters because a small set of unusually active organizations can dominate a pooled average. A median organization above 50% suggests the change reaches further into the customer base we observe. It does not mean the same organizations all crossed that threshold. It also does not mean AI wrote 57% of every codebase.
Our view
The next useful discussion starts after the tools are rolled out. Leaders need to understand which work AI helps complete, where that work gets stuck and whether assistance is spreading through the team.
What to do
Track adoption alongside AI-assisted output, matched cost per output and review flow. Before you read anything into low observed usage, ask which repositories and workflows still have missing telemetry.
Data behind this chapter
Observed AI adoption reached 82%
Percent of report engineers with observed AI activity
Quarter
Percent
Observed engineers
Total engineers
Q2 2026
76
16,245
21,409
Q3 2026
82
18,598
22,623
Population: Engineers with merged output in the quarter.
The sample changes from quarter to quarter. Detection depends on connected AI telemetry, so activity that is missing from telemetry can be recorded as human-authored.
AI-assisted output became common across organizations
Percent of merged output value attributed to AI-assisted PRs
Month
Pooled (%)
Median organization (%)
Apr 2026
41
21
Jun 2026
52
37
Sep 2026
57
53
Population: The pooled output share and the median organization share are separate aggregations.
A PR’s whole output is attributed when a detected AI contributor touches it. These are not percentages of code written by AI. Organization membership can change.
Higher-output bands also show deeper AI assistance
AI-assisted share of merged output value
Quarter
Below median (%)
Above median (%)
Q2 2026
40
49
Q3 2026
51
60
Population: Engineer output bands, drawn within each quarter.
This is an association. It does not show a causal AI lift. The Q2 above-median band excludes the 99th percentile, and the Q3 report text does not restate that exclusion. The bands are not fixed groups of workers.
AI-assisted output uses whole-PR attribution. The monthly pooled share and the organization-median share are different aggregations. No causal AI contribution is estimated.
The acceleration is visible beyond platform growth.
More organizations joining Weave cannot explain the entire output curve.
Between June and September, the published Index rose from 69.72 to 130.71 estimated expert output hours per employee. That is an 87% increase in three months. September alone was up 25.6% on August.
The Index is a blended average, so its population can change. We also examine organizations present in every quarter from Q3 2025 through Q3 2026. Within that fixed set, the median output change was 3.1×. That holds organization membership constant. It does not hold every engineer or every project constant.
The gains are uneven. In the same five-quarter cohort, 14.4% of organizations saw output decline. In the Q3 size benchmarks, p90 organizations produced roughly 2.5× to 6.8× the median within their size bucket. At 101+ engineers, the median was 46 output units per engineer per month and p90 was 143.
Our view
Engineering capacity is changing quickly, but an industry-wide average makes a weak operating target. The useful question is why a comparable team can turn the same broad technology into substantially more delivered work.
What to do
Compare yourself with organizations of a similar engineering size. Then inspect your own trend using stable repository coverage, work mix and quality measures.
Data behind this chapter
Engineering output accelerated through Q3
Estimated expert output hours per employee per month
Month
Output
Jan 2026
43.59
Feb 2026
49.66
Mar 2026
56.01
Apr 2026
56.4
May 2026
62.57
Jun 2026
69.72
Jul 2026
89.1
Aug 2026
104.06
Sep 2026
130.71
Population: Employees with AI spend or merged output in the month.
These are blended platform averages. A unit estimates the expert work represented by a merged PR. It is not a clock hour worked.
Growth remains strong with organizations held fixed
Multiple and percent, displayed separately
Measure
Value
Unit
Median organization output change
3.1
multiple
Organizations with declining output
14.4
percent
Population: Organizations active in every quarter from Q3 2025 through Q3 2026.
A five-quarter fixed set of organizations. The Q3 cohort count is not published. Engineers within each organization can still change. No quarterly path between the endpoints is published, so none should be drawn.
The output frontier sits far above the median at every size
Monthly estimated expert output hours per engineer
Team size
Q2 median
Q3 median
Q2 p90
Q3 p90
0–10 engineers
55
73
327
499
11–20 engineers
62
116
206
459
21–50 engineers
44
76
100
193
51–100 engineers
44
63
76
163
101+ engineers
29
46
74
143
Population: Organization-level median and p90, grouped by active engineering headcount.
Q2 and Q3 are cross-sectional groups. Bucket counts and standard errors are not published. The table shows dispersion. It does not show a causal effect of company size.
Output represents the estimated expert effort embodied in merged work. It is not time spent, realized business value or a controlled estimate of AI lift.
The output gains are real.So is the rising AI bill.
A bigger AI budget needs an output denominator.
From June to September, recorded AI spend per employee rose from $230.96 to $575.24, or about 2.5×. Output rose about 1.9× over the same endpoints. Q3 also introduced newly recorded cost types, so part of the spend increase comes from better coverage.
The matched spend-and-output series is a stricter way to look at unit economics. It includes only engineers with both attributed AI spend and merged output observed in the same month. On that basis, AI cost per output unit rose from $2.84 in January to $11.02 in August, then eased to $9.39 in September.
Token intensity is moving too. The report records 0.31 million tokens per output unit in January and 6.8 million in September. The source describes that as a 21× increase, calculated at its underlying precision. Expanded telemetry contributes to the observed change. More tokens can reflect longer contexts, more attempts or broader measurement. They do not directly measure more useful work.
September is worth watching. Output increased 25.6% while spend increased 10.7%, and matched cost per output fell 14.8%. One month does not establish a lasting efficiency trend. It does show why spend alone makes an incomplete dashboard: a larger bill can come with better unit economics, and a smaller bill can come with less output.
Usage is also highly concentrated. The top 1% of engineers ranked by output recorded median daily AI spend of $145 in Q3, against $4 in the bottom half. The source reports that spend for the top group rose 4.2× from Q2, against 1.8× for the bottom half. This is an association. It is no reason to give every engineer the same budget increase.
Our view
Manage AI as an input to engineering capacity. The objective is a relationship between spend, completed work and quality that you can defend, with enough freedom for high-value workflows to use more compute.
What to do
Give finance a matched cost-per-output trend and a note on coverage. Investigate expensive workflows at the session level before you impose a uniform spending ceiling.
Data behind this chapter
Spending grew faster across Q3; output grew faster in September
Index, June 2026 = 100
Month
Output
Spend
Jun 2026
100.0
100.0
Jul 2026
127.8
168.57
Aug 2026
149.25
225.08
Sep 2026
187.48
249.06
Population: The same published all-observed-employee Index denominator for each monthly series.
The population changes over time. Q3 spend adds cost types first recorded that quarter. An indexed comparison does not estimate causal AI return.
Population: Only engineers with both attributed AI spend and merged output in the same month.
Total observed spend divided by total observed output in this matched-signal population. The cohort expands from 2,800 in January to 9,151 in September. Unattributed organization-level spend is excluded.
Population: Observed token telemetry divided by shipped output.
Only these selected, rounded points are published. The months between them are not, and should not be filled in. Telemetry coverage expanded over the period. Token volume is not the same as spend, and a cache share measured only on Weave Router traffic cannot decompose the full platform.
The blended Index and the matched cost-per-output series cover different populations. The matched cohort was 2,800 engineers in January and 9,151 in September. This is not a financial ROI estimate.
Agent completion is only one step toward shipping.
The agent operating problem is conversion: turning activity into accepted software.
The Q3 split of session spend makes the distinction plain. Of every $100 in the analyzed spend, $23 was attached to sessions classified as completed with merged code. Another $17 was attached to merged code without a completion signal. In total, $40 was linked to some merged code.
A further $31 was attached to sessions classified as completed but not merged, $17 had no classified outcome and $12 was abandoned or blocked. These categories describe traced spend. They are not a count of tasks, and unmerged work can include research, experiments, rejected approaches or work still waiting for review.
The Q3 cut excludes one organization that accounted for 77% of traced session spend. That exclusion is material. The view is useful for understanding the rest of the observed spend. It is not a universal conversion rate for all Weave customers or for all AI engineering.
The practical distinction remains: a session can finish without producing an accepted change. A useful agent program has to follow the work past the last agent message, through completion, merge, review and deployment.
Our view
Teams should manage agent conversion as closely as they manage agent usage. The next gains may come from better task boundaries, review readiness, repository context and handoffs, even when the underlying model stays the same.
What to do
Review a sample of completed-but-unmerged sessions each week. Separate useful non-code research from rejected code, from work waiting for review and from work that never reached a clear owner.
Data behind this chapter
A completed agent session and merged software are different outcomes
Percent of traced session spend
Quarter
Completed and merged
Merged without completion signal
Completed, unmerged
Unclassified
Abandoned or blocked
Q2 2026
26
29
20
17
8
Q3 2026
23
17
31
17
12
Population: Spend-weighted session classifications. They are not a share of sessions or of engineers.
Q3 excludes one organization responsible for 77% of traced session spend. The remaining $17 per $100 linked to merged code lacks a completion signal. Unmerged work can include research, experiments or work still in progress.
The $23 completed-and-merged figure is carried over from the source unchanged. The remaining $77 should not be described as waste, and the additional $17 linked to merged code should not be ignored.
Codified workflows are associated with better completion.
How an agent is directed deserves the same attention as which model it uses.
In Q3, sessions that invoked a codified skill completed 45.3% of the time. Sessions without a skill completed 38.0% of the time, a gap of 7.3 percentage points. In Q2 the corresponding rates were 41.2% and 35.9%.
The comparison is specific. A codified skill is a reusable workflow invoked during a session. These figures do not compare high-skill and low-skill engineers. They also do not measure the newer AI Skills dimensions such as planning, verification or leadership.
Sessions that invoked a skill also added more lines on average: 630 against 51 in Q3. That is useful context about the kinds of work involved. It is also a warning against reading the completion gap as a clean experiment. Skills are often selected for larger, repeatable workflows, and lines added measure volume.
Our view
Repeatable work should become an explicit workflow that people can inspect, test and improve. When a task recurs, a prompt that worked should become shared operating knowledge.
What to do
Start with one frequent workflow. Define its inputs, checks and completion criteria, then compare outcomes before you expand it. Keep an audit trail from the skill invocation to the resulting PR.
Data behind this chapter
Codified workflows are associated with more completed sessions
Percent of agent sessions classified as completed
Quarter
With skill
Without skill
Q2 2026
41.2
35.9
Q3 2026
45.3
38.0
Population: Sessions that invoked a codified skill, against sessions that did not.
An unadjusted association. Skills are selected for particular workflows. These groups are not high- and low-skill engineers, and they are unrelated to current AI Skills V2 score quintiles.
Skill-invoking sessions also contain larger changes
Mean lines added per session
Quarter
With skill, mean lines
Without skill, mean lines
Q2 2026
744
138
Q3 2026
630
51
Population: Sessions that invoked a codified skill, against sessions that did not.
Lines are a volume metric. Codified skills are often designed for high-volume workflows. The 12.4× line ratio in Q3 is not a productivity effect or a causal return.
The association is unadjusted, and session counts and confidence intervals are not published. The result should not be labeled a causal completion lift or a gap in engineer ability.
Repository hygiene and context are engineering capacity decisions.
Q2 telemetry recorded about 3.3 million agent tool calls and roughly 81,000 failures. The largest named failure category was Not found: about 33,500 calls and 41% of failures. Timeouts, size limits, permissions, invalid arguments and version-control conflicts also appeared. Another 41% fell into an unclassified Other bucket.
This is a different reliability problem from whether a model can produce valid tool syntax. In September Weave Router traffic, only seven malformed blocks were observed among more than 1.46 million tool-use blocks. Valid syntax does not guarantee that the referenced file exists, that the environment is ready or that the task is making progress.
The Router also recorded 1,754 spiral-detector firings and 2,425 broader struggle firings in September. Detectors can fire more than once in a session, and the no-progress detector was introduced in late August. Use these counts to support investigation. They should not be drawn as a comparable platform-wide failure-rate trend.
Our view
Waiting for a stronger model is an incomplete reliability strategy. Teams can improve what the agent sees and what it can do: current instructions, reliable tests, discoverable paths, explicit permissions, clear task boundaries and a way to detect repeated failure.
What to do
Classify failures by cause and inspect recurring loops. Fix environment and context problems once, at the repository level, then measure whether the same failure comes back.
Data behind this chapter
Repository and environment failures deserve first-class attention
Approximate failed tool calls
Category
Approximate failed calls
Not found
33,500
Other unclassified
33,300
Timeout
2,600
Size limit
2,500
Permission denied
2,200
Invalid args
1,500
VCS conflict
1,300
Rate limit
1,200
Network
1,100
Syntax error
800
Server error
600
Interrupted
300
Population: About 81,000 failed calls among 3.3 million observed Q2 tool calls.
Category counts are rounded. 41% of failures are Not found, and another 41% are unclassified Other. A tool failure is different from a task failure or malformed tool syntax.
Q2 platform tool failures, Q3 Router detector events and malformed tool blocks are separate populations and outcomes. No causal improvement is claimed between them.
More output still needs a trustworthy acceptance process.
The report shows no broad review-time blowout. It does not close the quality question.
Median review turnaround was about 52 minutes in Q3, close to 51 minutes in Q2. Among teams above ten engineers, median review rounds stayed in a narrow 1.2–1.4 range. The revert rate for the latest month was 0.64%, compared with 0.98% in June.
Those are encouraging operating signals to see next to a steep output curve. They are also narrow measures. Recent reviews and reverts are floors, because later events have not arrived. A revert is only one way a change can prove problematic. It does not capture every production defect, security issue or maintainability cost.
Recorded AI-signed reviews accounted for 1.8% of review events in Q3, up from 1.1% in Q2 and 0.07% in Q4 2025. The figure describes signatures, so a human review that uses AI assistance may not appear as an AI review. It still makes the acceptance layer worth examining separately from code generation.
Our view
The engineering advantage will depend on how confidently a team can accept more work. Review, automated checks, production signals and clear ownership need to scale with agent output.
What to do
Pair output targets with review age, change-failure evidence and rework. Before you treat speed as the goal, check whether fast reviews reflect good preparation or weak scrutiny.
Data behind this chapter
Review turnaround stayed near 52 minutes
Median elapsed hours; rounded minutes
Quarter
Hours
Minutes
Q2 2026
0.85
51
Q3 2026
0.87
52
Population: Completed reviews with explicit review-request events. Weekends are excluded and nights are included.
The latest observations are floors. Teams without explicit request events are excluded. This measures one review interval. It does not measure end-to-end delivery time.
Population: Eligible merged PRs, with bot and noise exclusions.
The latest month is a floor because later reverts have not arrived. Reverts alone cannot establish overall quality, AI causality, production defect rate or long-term maintainability.
Review timing requires an explicit request and a completed review, excludes weekends and includes nights. No causal comparison of AI and human code quality is established.
A faster engineering system should be judged by the work it makes possible.
In Q3, feature work represented 56% of pooled output, keep-the-lights-on (KTLO) work 21% and bug work 23%. In Q2 the shares were 63%, 20% and 17%. Against Q3 2025, the larger change is a decline in KTLO share from 35% to 21%.
The rise in bug-work share deserves attention. It is not evidence that AI introduced more defects. Teams may be paying down an old backlog, changing priorities, classifying work differently or joining the sample with a different mix. The report measures the work that shipped. It does not measure the cause of every bug.
This is why capacity and allocation belong in the same conversation. A team can produce more output and still spend too little of it on the work that matters to its customers. The trends in feature, maintenance and bug work give leaders a way to ask where the new capacity is going.
Our view
Success with AI should include an explicit decision about where the additional engineering capacity goes. Backlog clearance, reliability work and customer features can all be valuable. The organization should be able to explain the choice.
What to do
Set a work-allocation goal alongside the output goal. Tie major changes in work mix to actual repositories, customer priorities and production evidence.
Data behind this chapter
More output still needs a work-allocation decision
Percent of pooled engineering output value
Quarter
Feature
KTLO
Bug
Q3 2025
54
35
11
Q2 2026
63
20
17
Q3 2026
56
21
23
Population: Merged output classified into feature, keep-the-lights-on and bug work.
A larger bug-work share is not a measured defect-introduction rate. It can reflect changed priorities, backlog clearance, classification and sample mix.
PR and line counts describe volume, which is only part of the result
Merged PR count and percent of merged lines, displayed separately
Month
Merged PRs, restated
AI share of merged lines, excluding generated (%)
Jun 2026
511,000
67
Sep 2026
650,000
76
Population: Pooled platform PR volume and non-generated merged lines.
The June figure of 511,000 is the restated Q3 pull. The original Q2 figure was 459,000. AI line attribution is a separate metric from whole-PR output attribution. The September AI share including generated files is 43%.
Model access does not explain the full distance between teams.
In Q3, median monthly AI spend per AI-active engineer peaked at $618 in organizations with 21–50 engineers. It fell to $433 at 51–100 and $301 at 101+. Organizations with 11–20 engineers recorded $602, and the smallest bucket recorded $410.
At the same time, the top decile of engineers produced 42–48% of output in organizations above ten engineers. Expertise remained concentrated too. Half of measured expertise sat with 21% of engineers in the 11–20 bucket and with 13% in the 101+ bucket. The source does not establish that spend or expertise concentration causes the output gap.
A spotlight on PE-backed companies shows why segment context matters. AI adoption was already 92% in Q2 and reached 93% in Q3, while AI-assisted output rose from 52% to 77%. Median recorded spend was $357 in Q3, or $261 when restricted to the cost types measured in Q2, against $96 in Q2. Adoption barely moved. Depth and spending moved substantially.
Our view
The next operating advantage depends on spreading reusable context and effective workflows through the organization. Buying a tool is easy to scale. Making expertise available at the point of work takes deliberate engineering.
What to do
Use benchmarks that fit your size. Inspect where expertise and agent usage are concentrated, and make high-performing workflows available to adjacent teams. Compare segment definitions before you borrow another organization’s target.
Data behind this chapter
Mid-sized engineering organizations record the highest AI spend
Median monthly USD per AI-active engineer
Size
Q2 (USD)
Q3 (USD)
0–10
229
410
11–20
264
602
21–50
255
618
51–100
235
433
101+
166
301
Population: Organization-level median within each active-engineer size bucket.
The denominator here is AI-active engineers, unlike the all-observed-employee Index. New Q3 cost types affect quarter comparisons.
Percent of engineers holding half of measured expertise
Size
Q2 (%)
Q3 (%)
11–20
24
21
101+
14
13
Population: Organization-level expertise concentration in the stated size buckets.
Q2 expertise uses a trailing twelve-month window through July. Only the Q3 endpoints are published, and the missing middle buckets should not be interpolated. This is concentration of code expertise. It is not a forecast of attrition.
Percent, with adoption and output on distinct series
Quarter
Adoption (%)
AI output (%)
Q2 2026
92
52
Q3 2026
93
77
Population: Hand-classified, majority PE-owned organizations with 50+ active engineers.
Q2 covered 17 organizations and 2,883 engineers. Q3 does not restate membership. This is an observational spotlight. It is not evidence that ownership causes AI outcomes.
Spend here is per AI-active engineer. It is not per all-observed employee. PE membership is hand-classified, with 17 organizations and 2,883 engineers stated in Q2 only. No ownership effect is estimated.
Context reuse and escalation belong in the economics of every agent workflow.
September Weave Router telemetry covered 1,624,315 requests across 52 installations. Of those, 1,329,835 were served by a different model than the client requested, about 82%. The resulting bill was approximately 44% lower than the requested models would have cost.
The largest savings came from small proprietary models. Two reroutes from Claude Sonnet 5, to Gemini 3.1 Flash-Lite and to GPT-5.6 Luna, accounted for about 45% of September savings. Open-weight models carried 1% of traffic and contributed 4% of savings. In this workload, a blanket move to open weights was not the best economic choice.
Cache state changes the answer. About 53% of roughly 234 billion input tokens were prompt-cache reads. Switching models can mean rebuilding a context that has already been paid for. In September, 33.1% of requests came from a session pin, and only about 1% of sessions with pinned turns were served by more than one model.
Weave Router treats that state explicitly. It considers the cost of rebuilding the cache before it switches, and it can escalate when the lower-cost model is not enough. The routing decision added a median 71 milliseconds, against a median upstream call of 4.6 seconds.
Our view
Teams should optimize the cost of finishing a coherent session. Price per token is one input. Cached context, failed attempts, model capability and escalation determine the bill the team actually pays.
What to do
Measure model choice, cache behavior, latency, spend and completion together. Hold the quality bar when you test savings, and evaluate repeated turns as a session instead of as isolated calls.
Data behind this chapter
Observed Router traffic grew more than sevenfold
Requests
Month
Requests
Jul 2026
222,946
Sep 2026
1,624,315
Population: All observed Router requests in the stated calendar month.
July and September are two months apart. They are not adjacent quarters. September has 52 installations, and its largest installation supplies 69% of requests. No July installation count is available.
Population: The September Router population: 1,624,315 requests, 52 installations and about 234 billion input tokens.
Savings compare the observed routing bill with the models originally requested. That is not a quality-matched counterfactual. The cache percentage is token-weighted, and the reroute share is request-weighted. This population is agent-heavy and distinct from the platform.
Router telemetry is a separate, agent-heavy population. One installation contributed 69% of September requests. Requested-model savings are not a measured quality-equivalent production counterfactual.
Routing can preserve frontier-level results at a lower cost.
Quality, cost and time should be evaluated together.
Weave Router and GPT-6 Astra were tested in the same Codex harness, with two attempts per task. On Terminal-Bench 4.0, the Router solved 62.1% of 66 tasks and Astra solved 60.6%. On SWE-Atlas Codebase QnA, the Router solved 62.1% of 124 tasks and Astra solved 66.1%.
The source treats those results as a quality tie. The Router and Astra confidence intervals overlap on both tests. The useful differences are cost and time. The Router used 52% of Astra’s cost per attempt on Terminal-Bench and 46% on SWE-Atlas, and its attempts were 2.2× and 2.5× faster.
These controlled results answer a narrower question than production telemetry does. They show what happened on two task sets under a stated harness. They do not establish that every routed production session has equal quality, or that pass@2 is the same as an engineer’s successful delivery rate.
Our view
The model decision should be an engineering experiment with a quality threshold and a budget for cost and time. A higher-priced default is a choice that can be tested. It does not have to be accepted without evidence.
What to do
Build a representative task set from your own work. Keep task success visible next to cost per attempt and elapsed time, then inspect where escalation is needed.
Data behind this chapter
Controlled tasks show frontier-level results at about half the cost
pass@2 percent; USD per attempt; minutes per attempt
Benchmark
System
Solved, pass@2 (%)
USD per attempt
Minutes per attempt
Terminal-Bench 4.0 (66 tasks)
Weave Router 2.0
62.1
5.22
20.5
Terminal-Bench 4.0 (66 tasks)
GPT-6 Astra
60.6
10.03
44.1
Terminal-Bench 4.0 (66 tasks)
OpenRouter Auto
25.0
5.52
36.0
SWE-Atlas Codebase QnA (124 tasks)
Weave Router 2.0
62.1
2.31
6.8
SWE-Atlas Codebase QnA (124 tasks)
GPT-6 Astra
66.1
5.04
16.7
SWE-Atlas Codebase QnA (124 tasks)
OpenRouter Auto
54.0
1.49
5.7
Population: 66 Terminal-Bench tasks and 124 SWE-Atlas tasks, two attempts per task, in the same Codex harness.
Router and Astra quality should be treated as a tie. Cost and time are per attempt. They are not per solved task. The Router and Astra runs are paired, and OpenRouter Auto is a separate run included for context. The Wilson intervals overlap. k=2 is not a Terminal-Bench leaderboard submission.
Pass@2 means a task counts as solved if either attempt succeeds. Costs and times are per attempt. The Router and Astra runs are paired, and OpenRouter Auto is separate contextual evidence. This is not a Terminal-Bench leaderboard submission.
A software factory needs a traceable path from intent to accepted change.
The current report already connects several parts of that path: observed AI assistance, agent-session outcomes, merged output, review, spend, failures and model routing. The next step is to make the joins between those records as explicit as the individual metrics.
The frontier questions are concrete. How much human intervention did a successful change require? How many agent attempts came before the accepted PR? Which skill, context package or model transition changed the outcome? How much review and rework did the change need after the agent finished? This report does not yet contain a defensible population-wide rate of fully autonomous delivery.
We think the engineering organizations that improve fastest will run those workflow records as a continuous feedback system. They will see where work stalls, change one part of the process and verify whether more useful software reaches the acceptance gate for the same or lower effort.
Our view
The next engineering advantage is how you run agents. The durable work is to make tasks legible, context reusable, checks reliable, review effective and economics visible.
What to do
Choose one recurring workflow and connect its task, sessions, PR, reviews, spend and deployment evidence. Establish the baseline before you add a success claim.
This chapter is an editorial outlook and a measurement agenda. It is not observed market prevalence or a published benchmark result.
Weave outlook
What we expect next
These are Weave’s editorial forecasts for the next 6–12 months. Each ends with the evidence that would support it.
Agent conversion becomes a leadership metric
We expect leaders to ask what share of agent effort becomes accepted work, and what happens to the rest. Adoption, token volume and session count will remain useful inputs, but they will be insufficient evidence for expanding a budget. We would count this as supported when teams can report a stable, coverage-aware view from completion to merge and explain changes in the unmerged share.
Repository context becomes part of the operating system
Skills, instructions, reliable checks and retrievable codebase knowledge will become maintained team assets. We expect teams to measure the workflow those assets support instead of judging them by how polished the prompt looks. The supporting evidence would be repeated workflows that show fewer avoidable environment failures and more accepted changes after specific context improvements.
The unit of AI optimization expands from tokens to sessions
As agents run longer and carry more context, cached state and escalation will matter more to total cost. Teams that optimize each request in isolation will miss part of the bill and part of the failure path. Look for session-level experiments that preserve task quality while reducing total spend or elapsed time, including cache rebuild and retry cost.
Quality becomes a continuous acceptance system
We expect review, standards, automated checks and production feedback to become more closely connected to agent execution. The objective will be to accept more useful work with confidence as generation accelerates. This would be supported if higher accepted output coincides with stable or improving mature measures of rework, incidents and change failure, and not only with faster code generation.
Measurement agenda
The next questions to answer
We cannot report these measures yet, and no results are invented here. Each entry gives the definition, what measuring it would need and the evidence that exists today.
Cost per accepted change
Attributed AI spend across every linked attempt, divided by the output value of the accepted merged change. Measuring it needs stable links from task to session to PR, cost coverage, an explicit accepted outcome and a way to handle research-only tasks. The related current metric is matched cost per merged output. A complete task-level roll-up of attempts is not supplied.
Human intervention per successful workflow
Human correction prompts, interruptions and required approvals for workflows that meet a declared success condition. That calls for an intervention taxonomy with timestamps, human and agent identity, a success condition and matched task complexity. The historical 93% no-correction, no-interrupt statistic is not task success and cannot fill this metric.
Attempt-to-acceptance ratio
Agent attempts and failed trajectories attached to a finally accepted change. It depends on task identity across restarts, branches and sessions, on the final accepted PR and on retry boundaries. Detector firings exist, but distinct attempted tasks and accepted outcomes are not published.
Post-acceptance rework
Revisions, reversions or remediation linked to accepted agent work within a declared mature observation window. The inputs are change lineage, time windows, bot and noise exclusions and a comparable mix of human and AI tasks. Reverts are published as a limited proxy. Complete rework data and mature production failure data are not supplied.
Context return
The difference in accepted workflow outcomes after a specified skill or repository-context change, on comparable work. We would need versioned context and skill invocation, a matched workflow population, and cost, acceptance and quality outcomes. An association between skill invocation and completion exists. A causal context return is not estimated.
Cache-adjusted session economics
Whole-session spend and completion, including cache reads, rebuilds, model switches and escalation. It takes session-linked model transitions, token types, billing, latency and the final outcome. Router cache and routing aggregates exist, and so do controlled benchmarks. A quality-equivalent production session experiment is not supplied.
Methodology and boundaries
Population
The Q3 report covers 22,623 engineers with merged output. Q2 covered 21,409 engineers across 1,470 organizations. The Q3 organization count is not restated.
Output
For each merged PR, Weave estimates the work an expert engineer would need to complete it, using source-code meaning, complexity and impact. Output is a modeled work unit. It is not elapsed labor time or realized business value.
AI attribution
AI-assisted output assigns the whole PR’s output when the PR has a detected AI contribution. It must be distinguished from AI line share and from an estimate of the marginal work AI performed.
Index
Output and spend per employee include employees with either observed spend or merged output in the month. Matched cost per output requires both signals in that same month and uses total spend divided by total output.
Timing
The Q3 Index snapshot was pulled October 7, 2026. Nothing here is a new October or November population estimate.
The sample skews toward organizations that measure engineering and is likely ahead of industry-wide AI adoption.
Population, connected-tool coverage and cost coverage change over time. Blended averages, fixed organization cohorts and matched-signal ratios answer different questions.
Weave Router observations cover 52 agent-heavy installations, and the largest contributes 69% of September requests. They are distinct from platform-wide results.
The Q3 session-spend distribution excludes one organization responsible for 77% of traced spend.
The latest review and revert observations are floors. Quality requires additional mature production and rework evidence.
Historical skill invocation, AI Skills scoring and agent autonomy are distinct definitions. No fully autonomous-delivery prevalence is measured here.
All Weave views and forecasts are editorial interpretations. They are not causal estimates from a randomized experiment.
Prepared November 5, 2026 from historical published aggregates. It is not a fresh October or November production pull.