agent-forge Pipeline: multi-agent builds with hard gates

Multi-agent builds fail quietly: a requirement gets dropped, the reviewer says "looks good", the verifier takes the implementer's word.

In the public run that built an exactly-once payments ledger, the verify gate halted twice (coverage 68%, 11 lint findings) and the reviewer caught a concurrency bug the implementer's tests missed. Three commits you can read: github.com/Ces107/agent-forge.

What the $24 adds over the free gates: the three gates (Spec Trace Gate, Adversarial Review Gate, Verify Gate) are free on their own. This adds the five sub-agent definitions, the orchestration that runs them in order, bounded fix loops that never lower the bar, and an optional tamper-evident record of what each stage produced.

The flow: you write the spec with your user (acceptance criteria as plain, testable sentences) → architect → planner → implementer → adversarial reviewer → verifier. Each is a sub-agent with restricted tools and its own model tier (opus / sonnet / haiku).

• Traceability gate: every acceptance criterion maps to a task and to a test that really exists, so no requirement is silently dropped.
• Review gate: findings count only with a failing test.
• Verify gate: tests, coverage, lint, types and security, each with its exit code.
• Fix loops: at most three rounds, and never by deleting tests, lowering thresholds or adding ignore rules.

In Cursor, Codex CLI and Copilot it runs stage by stage without sub-agents.

Launch price $24 until 3 November 2026, then $29.