Most guides survey the field. This one is the contract rubric.

An agent is in production when four artifacts live on main in your repo: the orchestration definition (LangGraph, Pydantic AI, or a custom DAG), an eval harness (ragas plus a case-specific rubric) running in your GitHub Actions and failing the build on regression, a failure playbook naming the fallback model and the audit-log schema, and an on-call runbook keyed to your rotation. Anything short of those four is a demo, not a production agent. The rubric thresholds were written on the week-0 scoping call and the week-2 gate is scored against them.

The named senior engineer is in your repo on day 1 with seat, CI, and branch access, and a reviewable PR lands by end of day 7. The PR is not a refactor or a discovery doc: it is the first piece of the happy-path agent executing end-to-end on a slice of your real input data, with an initial eval case checked in. If the PR does not land in seven days, weekly billing pauses until it does. The clause is in the MSA, not on a slide.

By end of week 2, a running prototype executes the happy path end-to-end and scores against the rubric you wrote with the engineer in week 0 (a min rubric score, typically around 0.82 on a case-specific scale, plus ragas thresholds around 0.78 on faithfulness and answer_relevancy). If the rubric gate fails, you get a prorated refund for weeks 1 and 2 and the engagement ends. The clause is in the MSA. No PIAS engagement has hit this exit so far, across five shipped production systems; the clause is the alignment mechanism, not the expected outcome.

Four artifacts on main, no external dependencies on PIAS: the orchestration definition (for example graph.py for LangGraph, pipeline.py for a custom DAG, or an agents/ module for Pydantic AI), the eval harness (ragas plus the week-0 rubric) running in your GitHub Actions, a failure playbook that names the fallback model and the audit-log schema, and a runbook keyed to your on-call rotation. A 90-minute handoff session with the on-call team closes the engagement. You own all of it.

For 12 months after handoff, the same senior engineer is available for 2-hour paid consults at a capped rate. Not a retainer. Not a SaaS subscription. Not an account manager calling every quarter. You buy 2-hour blocks when you hit a regression, a new failure mode, or want a second pair of eyes on a roadmap change. Cap prevents ratcheting. Your team is the owner of the agent by design; the engineer is a named fallback, not a dependency.

Seven days forces environment access, repo conventions, CI permissions, and code review to happen in week 1 instead of week 4. That is where most AI pilots fail: the prototype works on a vendor laptop but cannot integrate with the client's monorepo, so integration is pushed to the end of the engagement and never finishes. We invert the order: integration is the first week, modeling work is weeks 2 to 5. The PR clause is the mechanical enforcement of that order.

Those guides describe a framework or survey the industry. They do not commit a specific consultancy to a specific rubric with specific exit clauses on a specific calendar. This guide is the contract rubric five named production agents were held to: Monetizy.ai (around 8K emails per day on Pydantic AI), Upstate Remedial (400K plus emails on LangGraph), OpenLaw (publicly released), PriceFox (industry-leading automated eval CI), and OpenArt (custom scene-graph DAG). You can hold PIAS to this rubric. You cannot hold a survey article to anything.

The engineer picks on the scoping call, with written rationale, and the rationale is checked in next to the orchestration file so a future engineer can re-evaluate. Across the five shipped systems we picked Pydantic AI once, LangGraph once, and a custom DAG three times. The framework is downstream of the problem shape, not an opinion we walk in with. If the framework you picked before hiring PIAS is the wrong fit for the problem, we will say so on the scoping call, with the reason, before you spend another sprint on it.

The eval harness runs in your GitHub Actions (or your CI system of choice) on every PR that touches the agent code. Ragas plus the case-specific rubric from week 0, with thresholds that fail the build on regression. Traces land in your OTEL collector and your LangSmith project, on your keys. No PIAS-hosted dashboard. No vendor-hosted eval you have to log into. When the engagement ends, the only thing that changes is the name on the commits.

Yes, and a chunk of our engagements start that way. Week 0 is a read of your existing code plus your existing traces: named failure modes with no fallback, nodes with no eval signal, transitions with no audit row. Week 1 is the first PR that fixes one of those. Week 2 is the rubric gate against the cases you actually care about. We do not rewrite working code. If the framework you picked is the wrong fit for the shape of the problem, we will say so before touching a line of your repo.