The Intelligence Layer for Software Factories
How Weave connects code output, AI costs, agent sessions, quality and delivery so engineering leaders can improve work across people and coding agents.
On this page
Weave is an intelligence layer for software factories: engineering organizations where people and coding agents produce software together. It connects the work inside pull requests with AI usage, costs, reviews and delivery, so leaders can decide which tools to expand, which workflows to change and which problems need a closer look.
That connection matters when agents can generate changes faster than a team can evaluate them. Weave estimates the expert effort represented by code changes, puts AI spending beside that output and helps inspect review, quality and delivery evidence. Leaders can investigate the underlying work with Wooly, then act through a tool choice, a model choice or a repository standard. The decision starts with engineering evidence rather than the volume of generated code.
The execution stack can stay in place. Your coding agents, repositories, development environments and CI/CD system perform the work. Weave helps the organization understand and improve how that system performs.
Coding capacity is only one part of a software factory
The companies publishing software-factory accounts use different operating models. In February 2026, Stripe reported more than 1,300 agent-produced, human-reviewed PRs merged per week. Stripe reused its existing developer tooling and environments, combined deterministic checks with agent loops and limited CI repair to one or two rounds before returning unresolved work to a person.
OpenAI's harness-engineering account described a greenfield internal product where increasing agent code throughput made human QA a bottleneck. StrongDM's factory account instead describes a model built around specifications and scenarios without human code review. Those are different choices about acceptance and control.
An existing engineering organization needs evidence for its own choices: what work agents handle well, where people still intervene, how much accepted work results and what it costs. Copying another company's PR volume or review policy does not answer those questions.
What the intelligence layer must connect
A provider console reports tokens and spending. A repository records PRs and reviews. An agent harness captures execution traces. Weave adds an effort-based output measure across people and agents, connects AI costs to attributed output and brings review, quality and delivery evidence into the investigation. A buyer can ask which tool produces substantial work for its cost, then inspect whether that work creates review or repair problems. Coverage determines which records can be connected.
Scroll horizontally if needed →
| Work stage | Who performs or owns it | Evidence Weave connects | Decision it supports |
|---|---|---|---|
| Agent execution | Coding clients, the agent harness and the engineer directing the task | Captured conversations, session behavior, usage and cost sources | Change task context, verification practice, tool or model |
| Code production | Engineers and agents working in the repository | Estimated expert effort, cross-file context and the output-score explanation | Distinguish a difficult migration from a routine generated update |
| Review and validation | Reviewers, tests and CI; the team owns its standards | Review quality, response time, cycles and Checks comments | Change review ownership or make a recurring expectation repeatable |
| Release | The existing release process and its owners | Matched deployment and delivery evidence, reverts and later code turnover | Investigate where accepted work waits or needs repair |
| Investment and improvement | Engineering leadership and the team | Tool cost per output, model cost and output per session, labeled peer comparisons, developer feedback and Wooly's source-linked answers | Change the allocation or operating practice, then inspect the result |
A leader should be able to move from a trend to its cost source, pull requests, review evidence or captured conversations. Coverage and permissions determine which records are available; a missing link should remain visible as a coverage gap.
Measure the work inside the pull request
Weave's Code Intelligence estimates the expert engineering effort represented by a code change. Silk 1 reasons about the change and its dependencies across files, then explains the score. Teams can inspect that reasoning and correct estimates through overrides and organization-aware calibration.
A difficult migration and a generated dependency update can each count as one PR. Their engineering content can be very different. A factory that rewards PR count alone can mistake a larger number of changes for more substantial work.
Weave's output unit is estimated expert effort. A score representing three expert hours is a model's estimate of the work, rather than a record of elapsed time or paid hours saved. The Code Intelligence page reports 3.3× lower average prediction error for Silk 1 than Weave's original model, with 88% of PRs within ±1h of expert estimate. These are Weave's model-evaluation figures; the output score and its explanation are what the team can inspect on its own changes.
Put AI spending beside output and quality
Token Intelligence connects AI usage and costs with engineering work. The AI report includes output, cost and bug ratio by connected coding tool. Its model comparison adds average cost per session, first-shot success, average AI output per session and AI code turnover.
Cost per AI output helps compare spending against a common work measure. The cost source still matters. Provider-billed amounts, configured subscriptions, list-price consumption estimates and client telemetry answer different financial questions. A heavy user on a flat-rate subscription can consume more list-price-equivalent tokens without creating the same increase in the invoice. For these flat-rate tools, a person-level cost view can retain full list-price consumption even where tool-level reporting uses seat cost plus overage. Weave labels the available cost evidence; confirm the basis and attribution coverage before comparing views.
For an external reference, the product can compare AI cost per AI output with other Weave customers by organization size or, where available, industry. Those are separate peer comparisons, with the population identified in the product. They are also separate from the public Index's monthly matched cost-and-output cohort.
Use those views to investigate a tool choice. Keep financial ROI, cost efficiency and delivery improvement as distinct conclusions, each supported by its own inputs.
Understand how people and agents work together
A software factory still depends on decisions: how a task is scoped, what context an agent receives, how its output is verified and when a person intervenes.
Weave's Agent Observability exposes captured coding sessions and their available work context. AI Skills v2 samples one eligible session per engineer per day and examines decisions such as planning, verification and context management, with rationale linked to the conversation. The sample and capture requirements matter when interpreting the assessment.
A session may show that an engineer supplied a useful specification but accepted the result without testing an important edge case. The next action can be a verification practice for that work, supported by the actual conversation. Keep the one-eligible-session-per-engineer-per-day sample and capture coverage visible when discussing a specific planning or verification decision.
Developer-experience surveys add the team's account of tooling friction, documentation, focus and collaboration. Compare that feedback with engineering records to investigate a problem. Neither a session assessment nor a survey response should silently become a universal score of an engineer's value.
Use Wooly, Router and Checks to investigate and change workflows
Wooly answers cross-cutting questions using the organization's engineering records and links its answers to sources. A leadership review can ask where AI-assisted output is growing, which costs entered the comparison and where review or delivery needs attention. The next step is to open the evidence and choose an action with an owner.
Weave provides tools for model selection and repository standards:
- Model selection: Weave Router selects a model for coding requests. Weave's Router 2.0 announcement reported 40% lower token spending on average. That figure is a Router aggregate, separate from an organization's measured engineering results. Per-request model selection is distinct from deciding which engineering task belongs to a person or an agent.
- Repository standards: Weave Checks evaluates team-written Markdown standards against PR diffs and comments on violations. After a person or agent pushes a fix, a later run can resolve its comments. Checks does not rewrite the code; merge protection and deployment policy remain separate team configuration.
The team can investigate a recurring problem, change a model choice or encode a specific standard, then return to the work to see whether the change helped. The practical software-factory feedback guide develops that operating workflow.
Customer evidence for measuring people and agents together
Telnyx's June 2026 account describes 150 engineers working alongside 1,400 bots, including repository-specific maintainer bots that review changes against a repository's conventions and history. In Weave's public company announcement, Telnyx CEO David Casem describes using Weave to measure teams and agents. Telnyx's bots and their scale describe its own operating model at publication; they are not a Weave-built system or a Weave-caused productivity percentage.
Weave's public customer page reports 19% higher output per engineer at Reducto. The page does not state that result's observation window. The figure uses Weave's output unit, so a buyer can ask how that unit is calculated and inspect the underlying work.
OpenZeppelin used Weave output data to find a problem in one project early. Daniel Carmo estimated that the intervention probably saved four weeks of engineering time. The team reviewed output per engineer, turnover, PR cycle time and bug ratio weekly, then agreed specific actions with managers. This is an attributed estimate for one project incident, not recurring four-week savings.
Best engineering intelligence platforms for software factories
Weave publishes this comparison. Competitor descriptions reflect each vendor’s homepage as reviewed on November 14, 2026.
Scroll horizontally if needed →
| Platform | Human and agent measurement | Routing, coaching and workflow capabilities |
|---|---|---|
| Weave | Estimated expert effort per PR, an AI attribution hierarchy, source-labelled cost, and agent sessions linked to merged code | Router per-request model selection, AI Skills coaching, and Wooly questions with MCP access |
| DX | AI code tracking by commit, PR, team, agent and repository; agent-effectiveness assessments; complexity-weighted throughput | AI enablement and developer experience programs |
| Jellyfish | AI usage, cost and capacity across people and agents, tied to delivery outcomes | Investment and capacity reporting, with assistant workflows over engineering data |
| LinearB | Software-factory context, natural-language metrics, and cost per effective PR in sample workflows | Agentic Studio, MCP, model and effort selection, and gitStream review and merge workflows |
| Swarmia | AI impact, PR flow, investment balance and developer surveys | Custom reports, natural-language analysis, MCP and team improvement workflows |
| Faros | Coding spend and shipped outcomes, connected to an engineering data model | Model routes informed by evaluations built from the organization's code history, with usage governance |
Evaluate the intelligence layer on your own evidence
Bring a cross-file change, a routine generated update, an AI invoice and a captured coding session to the evaluation. Ask the platform to show:
- The work: how it estimates output and explains why the two code changes differ.
- The cost: for the same period and covered work, show the cost source, the AI-attributed output used as the denominator, and missing tool, person or session coverage. Keep billed amounts, configured seats and list-price usage estimates separate.
- The quality: what the reviews, reverts, turnover and delivery evidence say about that work.
- The human decisions: which captured session supports a coaching recommendation and how it was sampled.
- The action: what the team can change, who approves it and which evidence will show whether it helped.
Book a Weave demo and inspect the code, costs and conversations behind one engineering decision.
See the work behind your AI spend
Compare code output, AI cost and quality, then inspect the work behind the trend.
Related guides
Best engineering analytics tools
Compare productivity, review quality, delivery metrics and AI cost reporting.
Best developer productivity platforms for engineering teams
Compare approaches to measuring developer output, experience and AI impact.
Engineering platforms for understanding repetitive work and engineering investment
Understand where engineering effort goes across features, bugs and maintenance.