Bonfire

The 8 Agents

Eight production agent roles, each with a structural prompt that defines its constraints and capabilities.

Bonfire ships 8 production agent roles. Each role has a structural prompt (markdown) that defines what the agent can and cannot do, and a toolset that enforces those constraints at the API level.

Agent Roster

#AgentRoleStructural constraint
1ScoutObserver. Enters unknown terrain, returns a map.Read-only. Never modifies files.
2KnightContract author. Writes failing tests.Never writes implementation code.
3WarriorImplementer. Receives a contract, writes code to pass it.Never modifies test files. Up to 3 attempts.
4ProverIndependent auditor. Re-runs tests independently.On failure, bounces to Warrior.
5SageSynthesizer. Reads 2-3 competing reports, produces a decision.Output is synthesis, not merge.
6BardPublisher. Packages completed work as a pull request.GitHub API interactions only.
7WizardFinal quality gate. Reviews the PR.On rejection, bounces to Warrior.
8HeraldAnnouncer. Writes the session summary.Read-only on code; writes documentation.

Where they live

Each agent's structural prompt lives in bonfire/agents/<name>/prompt.md. The prompt is the agent's soul — it defines behavior through narrative constraints, not code.

Agent colors are canonical across the framework and the website:

AgentColorHex
ScoutBlue#60A5FA
KnightPurple#A78BFA
WarriorRed#F87171
ProverOrange#FB923C
SageGreen#4ADE80
BardPink#EC4899
WizardPink#F472B6
HeraldTeal#2DD4BF

Roles in the standard_build pipeline

The 7-stage pipeline uses 7 of the 8 agents in sequence: Scout, Knight, Warrior, Prover, Bard, Wizard, Herald.

The Sage is not part of standard_build. Sage fires only in dual/triple workflows where multiple parallel scouts need synthesis.

Structural prompts vs. system prompts

A structural prompt is not a system prompt. System prompts are runtime instructions that can be overridden or ignored. Structural prompts are compiled into the agent's context via the prompt compilation system with explicit priority levels. The structural prompt's blocks cannot be dropped by truncation — they are the highest-priority content in the agent's context window.

This means agent behavior is deterministic within the priority system. A Warrior with full context and a Warrior with truncated context both obey the same structural constraints; the difference is how much situational context survives truncation.

All tiers, all agents

Every tier — including free — ships all 8 agent roles. Tier differentiation happens through model class and vault access, never through agent availability.

On this page