The Right Model for Each Coding Task
Weave co-founders Adam Cohen and Andrew Churchill explain how their engineering analytics platform evolved from measuring human productivity to optimizing coding agents. The conversation covers Weave Router, which automatically selects the best model for each coding task, cost per output as a better measure than raw token spend, an AI coach for engineers, and the design of autonomous software factories with automated validation.
Weave: Turning Software Companies into Software Factories
Weave builds models that measure engineering work, then uses those measurements to optimize how humans and coding agents work together. The company started in Y Combinator’s Winter 2025 batch with a narrower focus: measuring human engineering productivity. Lines of code and PR counts had been the standard proxies, and the founders describe them as “fancy fun metrics to try to make executives feel good.” Weave replaced those with models that read the code itself and infer what an engineer actually did.
From Human Productivity to Agent Optimization
The pivot came from timing. Cursor was just taking off during the batch, and engineers were suddenly writing most of their code with LLMs. Customers asked Weave to quantify Cursor’s effect on their productivity, which made it clear that the same measurement approach had to cover coding agents, not only the humans supervising them. Early sales calls were met with skepticism, so the team connected each prospect’s data and showed them the results. Because their customers were fast-growing startups already working at the frontier of agentic coding, Weave says it has consistently been about six months ahead of the typical enterprise adoption curve.
The questions engineering leaders bring to the product have shifted along a predictable arc. Teams first want to drive AI adoption, then enter a “token maxing” phase where spending is the goal, and finally reach an optimization phase focused on ROI, wasted tokens, and whether agents are producing useful work. Startups tend to sit in the token maxing phase, while later-stage and private equity-owned companies are further along the optimization curve.
Cost per Output Over Raw Token Spend
The central metric is cost per output, defined as dollars spent on AI per unit of human-equivalent work, measured in hours. Raw spend is misleading in both directions. A team spending $10,000 a month that produces 10,000 hours of work pays a dollar an hour, which is a strong deal. A team spending the same amount for 10 hours of work pays $1,000 an hour, which signals a wrong model choice or an agent loop burning tokens without productivity gains. Cost and output have to be read together to tell whether AI spending is paying off.
Weave Router and Model Selection
Weave Router sends each prompt or task to the model best suited to it, so engineers don’t have to decide between Opus, Astra, or DeepSeek for every request. The router plugs into CLI-based coding agents such as Claude Code, Codex, and open-code with one command. The team reports that, on recent benchmarks, the router reaches Astra-level performance at half the cost and twice the speed.
The main concern with any router is that cheap models might quietly degrade code quality. Weave answers this with its own measurements. Its internal team, which uses the router heavily, reports output that keeps accelerating, placement in the top 1% of teams by Weave’s own data, and no increase in bugs or quality issues reaching production.
The router is trained on hundreds of thousands of real coding agent sessions. Training signals include what happens downstream: whether the resulting PR is merged, whether the change holds up in production, and how it performs a month later. The team also runs controlled testing across models and gives the router a reward when a task is completed successfully. The result is a learned model of each model’s strengths rather than hand-written rules about which model suits which job.
The distinction from OpenRouter is that OpenRouter functions more as a gateway, giving developers access to many models and fallbacks while leaving the choice to them. Weave makes the choice automatically and focuses on the coding agent use case, where its auto-routing is designed to perform well.
Subscriptions, Open Source, and Cost Savings
For users already paying for premium Anthropic and OpenAI plans, Weave can log into multiple accounts and route across those subscriptions, using subsidized capacity as long as it lasts and mixing in open-source models where they perform well. The core idea is that a frontier model is not always needed. A subagent exploring a codebase to find where a function is implemented can run on a smaller model such as Haiku-class or Luna without affecting the final result, and it preserves the larger plan’s usage limits. The team says this effectively doubles what the same subscription budget can accomplish.
Open-source models are not at the frontier, but they are close enough that their cost-performance ratio is very strong. DeepSeek, Kimi, and GLM models have performed best in Weave’s internal benchmarks and receive a significant share of routed traffic for simpler tasks such as centering a button or implementing a CRUD backend. The team argues that for most hobby projects, open-source models go surprisingly far, while complex work like building an ML classification pipeline still belongs with frontier models.
Measuring Engineers and Coaching Agents
Weave’s dashboards often surface outliers. Engineers whose AI spend looks alarming may turn out to be the most productive people on the team, and the data gives organizations a blueprint of what effective usage looks like. Leaders can then examine individual prompts to see what the strongest engineers do and spread those practices across the organization.
Parallelization is the highest-leverage practice the team has observed. Running many agent sessions at once is not enough on its own. The hard part is building guardrails and frameworks so that agents do the right work and humans step in at the right moments. This is an evolving practice that shifts with each new model release. Sharing skills, or refactoring parts of a codebase into architectures that agents handle well, helps teams spread the approach.
Weave’s product includes a coach that analyzes every session and provides daily recommendations, such as pushing back when an agent is wrong, restarting sessions more often because context windows make long sessions increasingly expensive, and delegating tasks more effectively.
Software Factories and Validation
The next stage after parallel agents is a software factory, in which agents keep building while the engineering team sleeps. The main challenge is validation, because an agent can ship anything, and output is worthless if it is not correct. Weave compares factory output with human-built work on outcomes, bug counts, and customer reports, and uses that comparison to decide which work can move toward lights-off operation. A customer named Uncountable is building such a factory and measuring it against its human teams in this way.
Weave’s own factory runs coding agents locally and in the cloud, with detailed guidance and a token-efficient CLI that gives agents the context they need. Each PR is reviewed by several automated tools, with the team using four review systems, three of which are from YC companies, and finding that the investment is justified because each tool catches issues the others miss.
The deeper problem is architectural. Every PR can be locally logical while the combined system becomes globally incoherent, filled with one-off workarounds. Weave addresses this with a checks system that encodes the judgments an expert engineer would make in review, such as preventing over-abstraction, and runs those checks on every PR. Agents do not learn from repeated feedback the way engineers do, so these checks act as persistent back pressure, and the team says the results have been strong.
Model Routing and Advice for Engineers
The team expects most engineers to stop manually choosing models and rely on routers that learn each model’s strengths. The next step is a router that outperforms any single model because it combines the best capabilities across them.
For engineers who want to lead their teams in AI use, the advice is to keep track of what the code is actually doing. The strongest engineers understand the system two levels of abstraction below and above their current work. That standard still holds with AI. The default level of abstraction has simply moved up, and losing touch with the underlying system is what makes people vulnerable to the wave of slop.
Looking ahead, the team believes running an engineering team will become an exercise in token allocation. The most valuable people will be those who identify problems worth solving, have taste and judgment, and direct agent swarms effectively, rather than those who are best at algorithmic puzzles such as inverting a binary tree.