How it runs

Agent sessions run in four levels: coordinator (Fleetforge), supervisor, orchestrators and implementers. Questions escalate one level at a time, and a session hands its work to a fresh one before its context runs full.

That setup carried the first long run, in 09.2026: about four days across three projects, in two stretches. On the client system alone it landed 73 reviewed work items in about 56 active hours.

I pick the models with blind A/B tests on real work items, using hidden tests and blind cross-vendor judges. In 09.2026 Claude Opus 5.5 at low reasoning effort replaced GPT-6 Luna as implementer, at 2–8× the speed with equal results. GPT-6 Sol stayed as reviewer, so no model checks its own work.

Who checks what

I added the fidelity auditor after an AI silently cut a 491-line spec to 122 lines.

Pipeline walkthrough

The screenshots below show the pipeline's stages on real runs.

Public tracker IDs (POE-###) stay visible. Private ones are replaced with [TSK-000].

01
Chunk dependency graph: backend foundation → tests + 3 independent frontends
Planner splits an epic into chunks and marks what runs in parallel after the foundation.
02
Plan review finishes with 5 severity-classified findings + a graceful proceed prompt
Plan reviewed before any code: 5 findings, each with a concrete fix.
03
8 agents in parallel: Group 1 light reviews + Group 2 implementers
8 agents at once: Group 2 implementers plus light reviewers for Group 1.
04
5 reviewers, 34 findings, 0 critical, clean to merge
5 review agents, 34 findings, 0 critical. Derived from the open-source pr-review-toolkit.
05
A real regression caught: flash key silently unified across two flows
Fidelity auditor catches a silent UX regression in a refactor: 20/21 OK, 1 critical.

Knowledge base

Pipeforge's own design decisions go into a knowledge base, each with its rationale, rejected alternatives and source files. It exists so I stop re-explaining why something works the way it does.

Pipeforge's CLAUDE.md in an editor. KB-first rule: call queryKnowledge() or listKnowledgeConcepts() via MCP as the first action whenever a prompt names a skill, agent, manifest, gate, contract, route, pipeline type or worker, contains terms such as autonomy, dispatch, orchestrator or triage, asks how or why something works, or proposes editing plugin files. Code exploration fills gaps after the KB, not before. KB degradation heuristic: after the KB call on a single-concept question, check whether the KB alone gave a correct, complete answer, and if not, write or enrich the entry
Pipeforge's CLAUDE.md: agents check the KB first and fill gaps when it falls short.