The most comprehensive report on the future of software engineering.

How AI is reshaping software development, measured from real engineering work.

Read the report ↓
Prologue

The next engineering advantage
is how you run agents.

Engineering teams have never had this much capacity. The teams getting the most from it have worked out how to run agents: what to hand them, what context to give them and how to check what comes back.

Weave is trusted by 500+ engineering organizations, from global startups to Fortune 100 and Fortune 500 companies. We see their merged code, agent runs, reviews and compute costs in one place, which shows us changes that adoption surveys miss.

Between June and September 2026, output per employee rose 87%. By September, AI-assisted pull requests accounted for 57% of merged output value. Spend rose faster than output, 2.5× against 1.9×, and a separate session analysis found that $23 of every $100 spent on agent sessions ended in work that was both completed and merged. The capacity is real. A lot of it still stops short of production.

Weave’s view

Engineering teams will compete on their agent workflows: which tasks they hand over, the context they maintain, the checks they trust and the decisions that carry useful work through review into production.

This report follows that path in order: how much output has grown, what it costs, what happens inside agent sessions, how review is holding up, how results differ between organizations and how model choice is changing. The Q3 study covers 22,623 engineers with merged output, alongside 21 months of Weave Index history, findings from the prior quarter and a separate Weave Router sample. Each population answers a different question, and we say which one is in use as we go. One limit to state up front: nothing here measures how much software agents deliver on their own, end to end. We measure the stages around that question.

22,623engineers with merged output in Q3
21 monthsof output history in the Index
Q3 2026July 1 – September 30

Based on production data through September 2026.

01
Output and adoption

The baseline
has moved.

Complexity-weighted merged output is climbing, and AI assistance has become routine in more of the organizations we observe. That resets what a normal quarter looks like.

Average monthly output per employee rose 87% in three months, from 69.72 units in June to 130.71 in September. It grew every month: 28% in July, 17% in August and 26% in September. The Q3 monthly average was 72% above Q2.

If you are still planning around last year’s throughput, the plan is probably too small. The useful questions now are which priorities can move sooner, what is still holding work back and whether the extra output holds up after release. Those are questions about how agent work is directed and checked, which is where the rest of this report goes.

An output unit models roughly one hour of expert engineering work embodied in merged code. It is not hours worked or business value delivered. Index populations change over time. In a separate fixed set of organizations, median output rose 3.1× year over year, from Q3 2025 to Q3 2026.

82%Observed AI adoption in Q3

We observed AI activity for 18,598 of the 22,623 engineers with merged output, up from 76% adoption in Q2. Usage is the easy part to count. The next figure asks how much of the merged work it touches. Connected-tool coverage and quarterly populations both change between quarters.

Q3 report · Observed AI use ↗

From June to September, the AI-assisted share of pooled merged output value rose from 52% to 57%. The median organization moved further, from 37% to 53%, which narrowed the gap between the typical organization and the pooled figure from 15 points to four. AI-assisted work is spreading beyond the largest contributors: in the median organization, over half of September’s merged output value came from AI-assisted PRs. When that much of the merged work has an agent in it, how those agents are run stops being a side project.

A PR’s full output value counts as AI-assisted if a detected AI contributor touched it. That is different from AI-written lines or AI’s marginal contribution. The pooled share and the organization median are aggregated differently, and the samples change, so this does not track the same organizations crossing a threshold.

02
The economics of building

More capacity.
A different cost curve.

Agents do not run for free. Compute spend is rising with output, and for most of the quarter it rose faster. What matters is how the two move together as teams get better at running agent work.

Figure 03

Spend outpaced output, then the monthly pattern turned

June 2026 = 100 · Monthly per-employee averages

AI spendEngineering output
Loading chart…

From June to September, AI spend per employee grew 2.5× and output grew 1.9×. Spend led for the quarter as a whole, but September broke the pattern of July and August: output grew 26% and spend grew 11%. One month does not make a trend, and we would not hang a claim about AI returns on it. Sustained output growth with slower spend growth is the pattern that would justify expanding agent capacity with confidence.

These are indexed per-employee averages. They compare growth rates and are not a matched cost per output or a financial ROI. Cost categories first recorded in Q3 also add to the increase.

Average attributed AI spend reached $575.24 per employee in September, up from $230.96 in June. The growth slowed as it went: 69% in July, 34% in August and 11% in September. A flat per-seat allowance cannot show whether an expensive agent workflow deserves more funding or needs repair. A budget set by workflow, with an owner and a delivered result attached, can.

The Index includes employees with AI spend or merged output. The organization-size benchmarks later in the report count AI-active engineers, so compare dollar values within one population.

For engineers with both spend and merged output, AI cost per output unit fell 14.8% in a month, from $11.02 in August to $9.39 in September. It is still well above January’s $2.84.

A unit cost only helps if you can trace it back to the work. Did the tasks get harder? Did agents need more attempts? Did routing or context change the bill? Session histories answer those questions, and they point at what to change.

Cost per output is total attributed spend divided by total output for engineers with both signals in the month. Ignored PRs and spend that cannot be attributed to an individual are excluded. This population grew from 2,800 engineers in January to 9,151 in September.

Recorded tokens per output unit rose from 0.31 million in January to 4.6 million in May, 4.4 million in June and 6.8 million in September. Two things are happening at once. Agent work has become more compute-intensive: longer context, retries and extended runs all consume tokens. And more tools began reporting token telemetry during the year, which lifts the recorded figure on its own. We show four months because those are the points the published reports support. Tokens are an input. Judge them by what each run did and what it produced.

Only selected months are published, and the months between them are left empty. Expanded telemetry coverage contributes to the change. Token counts and dollar costs diverge with model and cache-read prices. The 53% cache-read figure later in the report applies only to Weave Router.

Compute use is becoming more uneven. Median daily spend for the top 1% of engineers by output rose from $34 in Q2 to $145 in Q3. For the bottom half it rose from $2 to $4. The reported gap in monthly output widened from 84× to 118×.

This is why we argue for budgeting by workflow. An agent workflow that reliably delivers can justify a lot of compute. One that is stuck in retries needs a different response. Tell the two apart before allocating the next dollar.

Ranking engineers by output guarantees an ordering by output. Neither the spread nor its link to spend shows that spending caused it. Daily spend medians and monthly output medians cannot be divided to estimate ROI.

What happened to the work the team paid an agent to do?

03
The agentic development lifecycle

The important gap
is between done and shipped.

An agent finishing its session is not the same as the task being done. The change still has to be verified, reviewed and accepted, and this chapter is about what happens in that gap.

In Q3, $23 of every $100 of traced session spend went to sessions that completed and merged. Another $17 merged without a completion signal, so $40 in total is linked to merged work. The share that completed but never merged rose from $20 to $31 per $100 between Q2 and Q3.

Unmerged does not mean wasted. A completed session might hold useful research, a rejected approach, code waiting for review or work nobody owns. Each of those calls for a different response, and calling them all failures hides where the process can improve.

We think converting agent work into accepted work is one of the largest opportunities engineering leaders have right now. Sample the sessions that completed but did not merge, find where they stopped and give the recurring handoffs an owner. Clearer task boundaries and reliable verification improve the next run without changing the model.

The Q3 analysis excludes one organization responsible for 77% of traced session spend. With it included, $13 completed and merged and $41 remained unclassified, much of it in sessions that were still open. Shares measure spend, not task counts. Unmerged spend covers several outcomes and is not a single measure of waste.

A codified skill is a workflow written down so an agent can repeat it. In Q3, sessions that invoked one completed 45.3% of the time, against 38.0% for sessions that did not. The gap widened from 5.3 percentage points in Q2 to 7.3 in Q3. The two groups were doing different work, though: skill sessions added 630 lines on average and the others added 51, so read the completion rates alongside what was attempted.

The practical step is to make recurring workflows explicit. Define the input, supply the repository knowledge, set the acceptance checks and test each version on comparable work. What the team learns then stays in the workflow and does not have to be rediscovered in every session.

These groups are unadjusted and their task and workflow mixes differ, so the gap is not an estimate of a skill’s causal effect. Session counts and confidence intervals are unpublished. Lines added measure volume. Historical Q2 and Q3 skill invocation is separate from current Weave AI Skills scoring.

Q2 telemetry recorded about 81,000 failures across 3.3 million tool calls. The largest named category was not found: roughly 33,500 calls, or 41% of failures, where the agent reached for a file or path that was not there. Unclassified Other accounted for another 41%. Missing paths and unusable environments add cost to every run. Keep instructions current, tests reliable and permissions clear, fix the causes that recur, then check that later runs stop hitting them.

Counts are rounded. Agents can recover from a failed tool call, so these are not counts of failed tasks. Q2 platform telemetry is a different population from the September Weave Router signals below.

Reliability has several layers

A well-formed call can still go nowhere

7malformed blocks among more than 1.46M tool-use blocks
1,754spiral-detector firings in September Weave Router traffic
2,425broader struggle-detector firings in the same traffic

In September’s separate Weave Router sample, malformed tool-use blocks were rare, but detectors still found retries and broader struggle. Valid syntax is not progress. A workflow needs a way to notice a stall, then recover or hand the task back to a person.

These signals have different denominators. Detectors can overlap and do not count distinct failed sessions. Weave Router sample: 52 installations, approximately 234,000 sessions and 1.62M requests. Source and definitions ↗

04
Review, quality and work allocation

The human system
has to keep up.

After an agent executes, people still decide what gets accepted. As output grows, the questions are whether review is still catching what it should, how changes behave after release and whether the extra output is going where the team wants it.

If faster output were coming at the cost of broken changes, the revert rate is where it would show. The share of merged PRs that were reverted was 0.98% in June and 0.64% in September. AI-signed reviews grew from 1.1% of review events in Q2 to 1.8% in Q3, still a small presence next to AI’s share of output. Review wait barely moved: 0.85 hours in Q2 and 0.87 in Q3, about 52 minutes.

So far review has absorbed the extra volume. To keep accepting more agent work with confidence, teams need pre-review checks they trust, clear acceptance standards and human judgment kept for the changes that need it. Measure review speed together with what review catches and how changes behave after release.

Review wait covers completed reviews with explicit requests, excluding weekends but including nights. Late reviews can raise the latest estimate, so it is a floor. AI-signed share counts review events, not PRs with any AI review, and human-signed reviews may also use AI.

Observed September revert rate0.64%

A caution on that number. Reverts lag the merge, so September’s rate is a floor and may rise as later reverts are recorded. Reverts are also only one signal. To judge AI-assisted code properly, look at incidents, escaped defects and rework once that data has matured.

Q3 report · Latest-month floor ↗

Feature work fell from 63% of measured output in Q2 to 56% in Q3. Bug work rose from 17% to 23%, and keeping the lights on rose from 20% to 21%. More capacity does not decide where it goes. Someone still has to choose between product progress, maintenance and correction, and that choice shows up in what agents are asked to do. Read throughput alongside how it is being allocated.

These categories describe the work performed. Bug-work share is not a rate of defects introduced, and feature share does not capture all innovation or realized business value.

05
Organization benchmarks

The gains are broad.
The results are uneven.

Teams with access to the same technology are reporting very different results. Comparing organizations of a similar size shows variation the platform average hides.

Median output rose from Q2 in every size bucket. In Q3 it ranged from 46 units per engineer in organizations with 101+ active engineers to 116 in the 11–20 group. Within that group, the 90th percentile reached 459 units, almost four times the median. Even among organizations of the same size, results vary a great deal.

A benchmark cannot say which practices caused the gap, but it tells you where to look. Compare similar workloads, then examine how tasks are scoped, what repository context agents get and what happens between execution and merge. The goal is a process you can repeat without giving up quality.

These are cross-sectional benchmarks whose membership and size buckets change, unlike the fixed cohort below. Bucket counts and uncertainty estimates are unpublished, and the largest buckets have fewer observations.

3.1×Median year-over-year output change

Among organizations present in all five quarters from Q3 2025 through Q3 2026, the median change in output was 3.1×. Yet 14.4% of them declined. Holding membership constant supports the growth finding and shows how far individual experiences diverge from it.

Q3 report · Constant-organization cohort ↗

In Q3, median monthly AI spend per AI-active engineer was $410 at 0–10 engineers, $602 at 11–20, $618 at 21–50, $433 at 51–100 and $301 at 101+. The middle bands spent the most. Budget policy, workload and how intensively teams use agents could all play a part. Look at those, and at what the spend delivered, before treating a peer’s number as a target.

This chart counts AI-active engineers. The earlier Index figures include employees with AI spend or merged output. The association does not establish an effect of organization size.

Organizational resilience

In teams above 10 engineers, the top decile produced 42–48% of Q3 output, up from 38–43% in Q2. The context, checks and agent workflows behind that output are worth sharing, so the gains reach the rest of the team.

Q3 organization benchmarks ↗
06
Models, routing and the execution layer

The workflow extends
beyond one model.

Models change faster than teams can rebuild their habits around them. The knowledge of how to run agent work has to outlive any one model, and someone has to decide which model serves each request, what context to reuse and when a hard step should escalate.

Opus 5 accounted for 71.4% of recorded September tokens in the Index, after earlier generations had held much of the workload earlier in the year. Workloads move to a new model within months. To evaluate one, keep the tasks stable, make the quality checks explicit and compare costs on equal terms. The context and practices that made agent work reliable should carry over.

The five named families have the highest average monthly token shares from January through September. Other combines the remaining families and the source’s unallocated Other category. Family names follow the Weave Index, which labels Claude Opus families as Opus. Tokens measure workload volume, not developer adoption, provider revenue or market share.

Figure 16September 2026

Model choice happens
inside the workflow

Weave Router served 1,329,835 of 1,624,315 September requests with a different model than the one requested.

82%

of requests rerouted

~44%

Lower billed cost than the
requested-model counterfactual

53%

Of input tokens
were cache reads

Weave Router sample: 52 installations and about 234,000 sessions in September. The largest installation accounts for 69% of requests, so read this as a small, agent-heavy population and not the platform as a whole.

In the separate Weave Router sample, about 82% of September requests were served by a different model than the one requested, and the bill came to approximately 44% less than the requested models would have cost. Cache reads made up 53% of input tokens, so cost depended on reused context as well as on model choice.

Routing is an engineering decision made inside the work, one request at a time. Use a cheaper model where it is good enough and escalate the harder steps. Context reuse changes the cost of either choice. Judge the whole session against a quality standard you have written down.

Weave Router is a separate population of 52 installations and approximately 234,000 sessions, and the largest installation accounts for 69% of requests. Savings against the requested model compare bills. They are not randomized or quality-equivalent production outcomes. Q3 AI Impact Report ↗

The economics of a coherent session

Context changes the routing decision

33.1%of September requests used an automatic session pin
~1%of sessions with pinned turns were served by more than one model
71 msmedian routing decision, versus a 4.6-second median upstream call

Switching models can mean rebuilding context you have already paid for. Session pins help preserve that continuity: about 1% of sessions with pinned turns used multiple models. Measure cost across the whole session, including cache rebuilds, retries and escalation.

September Weave Router telemetry only. Pin share counts requests, continuity counts sessions with pinned turns and cache-read share counts input tokens. These are separate measures. Q3 routing and cache analysis ↗

Controlled benchmarks put these tradeoffs to the test. On Terminal-Bench 4.0, Weave Router solved 62.1% of tasks at pass@2 against GPT-6 Astra’s 60.6%, at 52% of Astra’s cost per attempt and 2.2× the speed.

On SWE-Atlas Codebase QnA, Weave Router solved 62.1% against Astra’s 66.1%, at 46% of the cost and 2.5× the speed per attempt. The source treats quality as a tie on both suites because the confidence intervals overlap. Read the three together: quality, cost and time, under these test conditions.

In your own codebase, set the quality bar first, compare what it costs and how long it takes to reach it, then validate in production. Completion, accepted output and later rework will tell you whether a benchmark result holds for your agents.

Terminal-Bench 4.0 has 66 tasks and SWE-Atlas Codebase QnA has 124. Weave Router and Astra were run as a pair; Auto ran separately within three days. Pass@2 counts a success in either of two attempts, and cost and time are per attempt. Benchmarks do not measure production success rates. See the source for pricing and confidence intervals.

The Weave outlook

Running the factory
becomes the advantage.

We expect the next engineering advantage to come from the whole path a task takes, from the moment it is chosen to the moment it runs in production.

Output, AI assistance and compute use have all grown. An agent completing its session still leaves acceptance unresolved, and organizations of a similar size are getting different results. That is where leaders should invest: in how work is directed, checked and delivered.

We expect software factories to connect task context, agent execution, verification and review, and to feed accepted changes and their costs back into the next task. Human judgment stays essential. What changes is that the process becomes visible, and a visible process can be improved.

Choose the work→Supply context→Run & verify→Review & ship→Measure & improve
01

Agent workflows get owners

Recurring work will have defined inputs, acceptance tests and an accountable owner. Teams will refine each workflow from one run to the next. Progress will look like more comparable tasks accepted, fewer retries and less avoidable intervention.

02

Budgets follow the work that delivers

High-output engineers already spend very differently from the middle. We expect budget conversations to shift to which workflow deserves more funding, what extra work it should deliver and how quality changes as a result. That makes the return on intensive agent work easier to see.

03

Context becomes shared engineering infrastructure

Repository knowledge, reusable instructions and verification need maintaining as the code changes, like any other infrastructure. Their value shows when another engineer or another agent can repeat the workflow. With output this concentrated, sharing that knowledge matters all the more.

04

Acceptance becomes a continuous process

As generation speeds up, checks, review and production feedback need to connect directly to agent execution. A team should be able to say why a change was accepted and whether it held up. The signal to look for is more accepted output with stable or improving mature measures of rework, incidents and change failure.

The next measurement frontier

How much useful software can a team deliver autonomously, at a known cost and quality?

What we can measure today covers AI-assisted output, session outcomes and routing. Measuring software-factory delivery across a whole population needs more: linked task-to-release records, human intervention, verification, accepted output and full cost.

The measurements that would make the next edition more useful

This edition measures parts of the process. It does not measure a population-wide rate of autonomous delivery. To get there, follow a recurring task through every attempt, human correction, accepted change and post-release outcome. Connecting those records gives a team a baseline to improve on and gives researchers a unit they can compare.

MeasureWhat it would tell us
Task-to-release conversionHow often accepted tasks reach verified production releases, and where they stop.
Human intervention per delivered taskThe correction, review and supervision each delivered task requires.
Cost per accepted taskCompute, retries and human work per accepted result.
Verification coverage and escaped defectsWhether checks catch consequential failures as delivery accelerates.
Workflow repeatabilityWhether results hold across engineers, models, repositories and task difficulty.
Marginal compute returnThe output from additional spend on comparable work.

These six measures are research priorities for future editions.

If you do one thing after reading this, pick a single recurring task your team already gives to agents. Write down what done means, keep its context where the next run can find it and follow a few runs all the way to merge. You will find out quickly where the work stops. That is the place to start.

Sources and methodology

Behind the numbers.

This edition draws on three kinds of evidence: the published Q2 and Q3 production aggregates, the Weave Index snapshot pulled on October 7 and the controlled Weave Router benchmarks. Production observations end in September 2026. Every figure links to its source and states its period and unit.

Weave’s reach as a company is wider than the populations measured here. The Q3 study covers 22,623 engineers with merged output. The Index, the session analysis and the Weave Router telemetry each have their own inclusion rules. Company coverage ↗ · Customer context ↗

Platform report22,623 engineers

Engineers with merged output in Q3 2026. Organizations that already measure engineering are overrepresented, so observed adoption may run ahead of the wider industry.

Weave IndexMonthly snapshots

Output history from January 2025 through September 2026, pulled October 7. Monthly blended averages reflect changing platform membership.

Weave Router telemetry52 installations

September traffic only, about 1.62 million requests. The largest installation accounts for 69% of them.

Matched cost / outputBoth signals observed

Engineers with attributed spend and merged output. Ignored PRs and spend unattributed to individuals are excluded.

How output, AI attribution, and spend are defined

Weave models the expert work required for merged code using its meaning, complexity and impact. One output unit represents roughly one hour of expert work embodied in the change, not elapsed labor time or realized business value.

If a detected AI contributor touches a PR, its full output value is classified as AI-assisted. This measures output associated with observed assistance, not AI-written lines or AI’s marginal contribution.

Index per-employee averages include employees with AI spend or merged output. Cost per output requires both and divides total spend by total output within that matched population.

Important sample and comparison notes

Q3 session-spend outcomes exclude one organization with 77% of traced spend. Some unclassified sessions remained open at quarter end. Newly captured Q3 cost categories can inflate comparisons with Q2.

Blended trends reflect changing platform membership; fixed-cohort analysis holds organizations constant. Observed relationships among skills, spend and output do not establish causation.

Latest revert and review measures remain floors. Token intensity uses published observations and reflects expanded telemetry coverage. Team-size and work-mix comparisons are cross-sectional. Fixed-task benchmarks differ from production success rates. Model mix measures recorded tokens, not engineers, spend or market share. Historical skill invocation differs from current AI Skills scoring. The outlook and proposed measures express expectations and research priorities, not additional population findings.

Sources, definitions and downloadable data

Questions this report answers

How much did engineering output grow in Q3 2026?

Average monthly output per employee in the Weave Index rose 87% between June and September 2026, from 69.72 to 130.71 output units. The Q3 monthly average was 72% above Q2.

How much are engineering teams spending on AI?

Average attributed AI spend reached $575.24 per employee per month in September 2026, up from $230.96 in June. Over the quarter, spend grew 2.5× while output grew 1.9×.

How much merged code is AI-assisted?

In September 2026, AI-assisted pull requests accounted for 57% of pooled merged output value, and 53% in the median organization. AI activity was observed for 18,598 of 22,623 engineers with merged output.

How much AI agent spend ends up in merged code?

In Q3 2026, $23 of every $100 of traced agent session spend went to sessions that completed and merged. Another $17 merged without a completion signal, so $40 is linked to merged code. $31 completed without merging.

Is AI-assisted code being reverted more often?

The share of merged pull requests that were reverted was 0.64% in September 2026, down from 0.98% in June. September is a floor, because reverts can be recorded after the merge.

What is an output unit?

An output unit models roughly one hour of expert engineering work embodied in merged code, based on the meaning, complexity and impact of the change. It is not hours worked or business value delivered.

Where does the data come from?

From production data on the Weave platform: merged code, agent sessions, reviews and AI spend. The Q3 study covers 22,623 engineers with merged output between July 1 and September 30, 2026. It is not a survey.

Weave Research

Production data through September 2026.
Index snapshot pulled October 7, 2026.

↑