Agentic AI and build-first workflow

CAIRE's engineering execution model is an agentic PRD-to-PR pipeline. A human writes a product brief; specialised AI agents handle spec, tests, implementation, review, and verification; the human approves a screenshot. Compute is the bottleneck — not headcount.

From a sentence to shipped software. The human appears in two places only — writing the PRD, and approving the dossier screenshot.

The agentic workflow is not "AI assistance for engineers". It's the execution model. Four commitments make the workflow distinctive.

Product owners write PRDs. The pipeline takes the PRD from there — spec, tests, code, review, dossier, merge — without a human in the middle of any stage.

Every model call goes through an adapter. Routing is config, not code. When a better or cheaper model ships, rotation is a quarterly review — not a refactor.

Post-merge metrics ramp a feature 1% → 100%. Regression flips the flag back. The decision is mechanical — humans don't decide "ok, ramp this".

Cash balance, revenue, and burn are system inputs. The orchestrator refuses to spawn an expensive run if today's budget is exhausted. No human "tighten the belt" call.

Every architectural decision in the pipeline is graded by one question: does it make us ship more features per second, per dollar (and per token)? The metric is deliberately tiny in absolute value. What matters is the trajectory.

Wall-clock from PRD intake to merged PR. Squeezing this means parallelising stages, caching specs, removing human round-trips. Every shipped feature lands a row in .compound-state/agent-service.db with its elapsed time.

Total model spend across all eight stages, per merged PR. Cheaper providers, smaller models for cheap routing, batch APIs, prompt caching — every lever points back here. The orchestrator refuses runs whose projected cost would exceed today's budget.

Total input + output tokens across the pipeline, per merged PR. The cleaner the spec and the tighter the dossier contract, the fewer tokens the editor burns iterating. Tokens are a leading indicator of cost.

An optimization mathematician agent reads the throughput log every week and proposes routing changes — different model per role, different batch size, different cache strategy. The CPO/CTO agent ratifies. The ratchet only moves one way.

This page itself will move with the metric. New routing wins, new agent prompts, new dossier shapes — every improvement that nudges features-per-second-per-dollar lands here as a refresh.

Each stage produces a checked artefact. The next stage refuses to start without it. The pipeline is one continuous flow, not a checklist — every stage hands off a typed result.

Re-entry: verify and the reviewer loop may send work back to implement (max 3 cycles). Merge runs only after human Approve — not in the unattended walk.

The contract is sharp: no dossier, no merge . Every stage's output is a typed, persisted artefact that the next stage reads — and that a human, an audit, or a future agent can replay.

The agentic workflow doesn't make humans disappear — it makes them strategic. Each role has one or two narrow places to step in. Everything else is software.

Defines what + why Writes the PRD in wiki/plans/ with a success metric and explicit non-goals. Approves (or rejects) the dossier screenshot before merge. Sets the regression threshold that scale-or-kill watches post-merge. 💻

• Writes the PRD in wiki/plans/ with a success metric and explicit non-goals.
• Approves (or rejects) the dossier screenshot before merge.
• Sets the regression threshold that scale-or-kill watches post-merge.

Reviews, doesn't author Reads the auto-generated spec for drift from the PRD. Argues down P2 reviewer comments with a justification when the agent is wrong. Maintains the agent prompts and the model adapter — code about how code gets written. 🛠️

• Reads the auto-generated spec for drift from the PRD.
• Argues down P2 reviewer comments with a justification when the agent is wrong.
• Maintains the agent prompts and the model adapter — code about how code gets written.

Owns the substrate Operates Darwin (the orchestrator runtime), the launchd job slots, the merge queue. Watches the throughput log: cost per merged PR, features per second per token. Approves model-routing rotations from the optimization mathematician's weekly proposal. 🧪

• Operates Darwin (the orchestrator runtime), the launchd job slots, the merge queue.
• Watches the throughput log: cost per merged PR, features per second per token.
• Approves model-routing rotations from the optimization mathematician's weekly proposal.

Writes the rules, not the cases Curates the Gherkin patterns the Test-writer agent compiles from. Audits dossier summary.json for skipped scenarios or empty Playwright traces. Owns the "no dossier, no merge" gate — the only thing the orchestrator cannot skip. 📈

• Curates the Gherkin patterns the Test-writer agent compiles from.
• Audits dossier summary.json for skipped scenarios or empty Playwright traces.
• Owns the "no dossier, no merge" gate — the only thing the orchestrator cannot skip.

Watches the leverage Reads cost-per-merged-PR trending down month over month as the routing matrix tightens. Tracks features-per-second-per-token as the leverage ratio that doesn't depend on hiring. Ratifies quarterly model-routing decisions; does not pick models. 🔍

• Reads cost-per-merged-PR trending down month over month as the routing matrix tightens.
• Tracks features-per-second-per-token as the leverage ratio that doesn't depend on hiring.
• Ratifies quarterly model-routing decisions; does not pick models.

Audits the trail Inspects wiki/raw/dossiers/<feature>/ on a merged PR — full Playwright trace, screenshot, console log. Reads the PRD frontmatter to map a shipped feature back to the original brief. Verifies vendor-agnosticism by reading the routing config — no provider lock-in to inherit.

• Inspects wiki/raw/dossiers/<feature>/ on a merged PR — full Playwright trace, screenshot, console log.
• Reads the PRD frontmatter to map a shipped feature back to the original brief.
• Verifies vendor-agnosticism by reading the routing config — no provider lock-in to inherit.

The pipeline targets steady-state product engineering — the work that, in a traditional team, fills standups and sprints. Bigger architectural moves still get a human-led plan.

A new banner, a new page, a form variation. PRD names the acceptance scenarios; pipeline writes vitest + Playwright tests; Editor implements; Verifier captures the screenshot dossier.

PRD frames the bug as a failing scenario. Test-writer compiles it; Editor fixes; resolver-reviewer + perf-reviewer catch N+1 regressions before the PR opens.

Throughput row, KPI tile, dashboard chart. Spec names the data source; tests assert the shape; the dossier shows the metric rendering with realistic seed data.

Quarterly: the optimization mathematician proposes a routing change based on cost, pass rate, latency. CPO/CTO agent ratifies. One config edit; the adapter handles the rest.

Audit pass like the agentic-workflow audit itself: identify drift, fix the doc, run yarn wiki:lint , ship. Wiki-only PRs use the same eight stages with the verifier in light mode.

A hypothetical external data flow starts with a PRD. Tests cover the boundary, and the dossier records what was actually verified. The example describes the engineering method — not a shipped integration or product catalogue.

Autonomous loops without guardrails are how AI projects burn budgets and ship regressions. Every stage of the pipeline is fenced. The Editor agent sits at the centre, surrounded by mechanisms that can either slow it down, redirect it, or stop it entirely.

Eight independent guardrails. No single one prevents bugs alone; together they make autonomous shipping safe enough that the human's only required action is reading a screenshot.

The Editor cannot ship behaviour the spec didn't name. Drift between the PRD and the spec is itself a P2 finding for the verifier — and the spec is short enough to fit in every agent's context, so drift is always provable.

MAX_ITERATIONS = 8 on the editor's inner loop, plus a max of 3 re-entry cycles from review or Codex feedback. Hit the ceiling and the run surfaces failure to a human — it never grinds.

The optional PIPELINE_BUDGET_ENFORCEMENT flag (off in pilot, on once revenue is real) refuses further model calls when the per-run cost would exceed the cap. In pilot the human is the only PRD producer, so cost is implicitly bounded.

A green CI run is necessary but not sufficient. The dossier — Playwright trace, final screenshot, console log, machine-readable summary — proves the feature actually rendered. Reviewers can replay the trace; the human sees the screenshot.

Same pipeline, three ways to enter it. Pick the surface that matches the moment — a PRD in version control, a form on the Dashboard, a message on Telegram. Every surface produces the same dossier and the same merge decision.

Write wiki/plans/<feature>-YYYY-MM-DD.md , create an isolated worktree with ./scripts/git/worktree-add.sh , and the pipeline runs against that branch. Reviewer subagents lint the diff, the verifier captures the dossier, the merge queue lands the PR. Best for engineers shipping in version control.

Paste a PRD path, click start , watch the pipeline progress at localhost:3010 . The dossier viewer surfaces the screenshot, console log, and machine-readable summary inline. Approve / Reject is a button — no GitHub round-trip required. Best for product owners who want a UI, not a CLI.

Post a PRD link to interface-agent . The same backend runs the pipeline; the dossier screenshot posts back to the originating thread. Reply approve and it merges. The lightest possible surface — a notification and an image. Best for the founder reading on a phone between meetings.

The pipeline is built on three open patterns and one discipline. Nothing here is bespoke for the sake of bespoke.

One expensive reasoning pass produces the spec; many cheap edit passes implement against it. ~1/14th the cost of running every call on the reasoning model — the cost driver is the editor, not the architect.

Plain language is too imprecise to coordinate multiple agents. Acceptance scenarios in Gherkin are short enough to fit every agent's context, precise enough to compile to failing tests, and greppable.

A green CI run is necessary but not sufficient. The dossier — trace, screenshot, console log, machine-readable summary — proves the feature actually rendered, in a real browser, in the state the spec named.

External review bots (Codex, CodeRabbit) post comments after every push. The pipeline polls them, treats P1/P2 as failed tests, and re-enters the editor automatically. The discipline is non-optional.