// about
One developer. Eight agents.
Six structural decisions.
Every agent framework dispatches. Only Bonfire learns, competes with itself, and feeds its own fire.
Bonfire is an AI agent orchestration framework built on the Claude Agent SDK. It started as a personal tool — one developer solving the same production problems every other framework punts to the application layer.
The framework ships six structural decisions that compound: a pipeline engine where workflows are data, priority-based prompt compilation that no competitor ships, dual/triple workflows for structured disagreement, TDD enforced at the pipeline level, a local-first knowledge graph, and per-stage fuel tracking.
45+ development sessions. 728 test files. 8 production agents. Built in public, receipts on every claim.
// the eight
Agent Roster
Scout
Observer
Enter unknown terrain, return a map. Read-only. Never modify files.
Knight
Contract Author
Define what correct means before anyone writes a line of implementation.
Warrior
Implementer
Receive a contract and make it pass with the minimum correct implementation.
Prover
Auditor
Re-run tests independently. Trust nothing the Warrior claims.
Sage
Synthesizer
Read competing reports. Produce a decision, not a merge. One answer per axis.
Bard
Publisher
Package completed work as a pull request. The PR is the artifact.
Wizard
Quality Gate
Review the PR. On rejection, bounce to Warrior. The Wizard's word is final.
Herald
Announcer
Write the session summary. Surface what shipped.
All 8 agents ship on every tier, including free. Tier differentiation happens through model class and vault access, never through agent availability.
// philosophy
What we believe
- Free is the product, not the demo.
- The free tier ships all eight agents, all seven pipeline stages, full priority truncation, and the full gamification loop. Monetization sits above via plugins and tier-gated services, never via engine feature removal.
- Every claim carries a receipt.
- No adjective when a number fits. 728 test files, not “comprehensive tests.” 8 agents, not “a powerful multi-agent system.” If a claim can't cite source, it doesn't ship.
- Admit what doesn't work.
- The limitations page is mandatory. What Bonfire can't do is as load-bearing as what it can.
- Competing solutions, not task-splitting.
- Dual and triple workflows dispatch multiple agents on the same problem. The Sage synthesizes the winner. One of them is always worse — the Sage says so and documents the tradeoff.
- The Vault earns its keep or it doesn't ship.
- Vault rollout is gated by founder evaluation, not a release calendar. No “coming Q3” promises. The dogfood phase runs until the experience is real, then it lights up.
// receipts
Read the spec
The full v1 architecture spec is public. Every claim on this site traces to a source file, a test count, or a session handoff. The spec is the contract between the website's copy and the framework's reality.
Read the architecture docs