Skip to content

Agents Overview

Agents Overview

claude-forge uses 10 specialist agents, each handling a single phase of the pipeline. Every agent runs in an isolated context window with its own system prompt defined in a .md file.

Agent Roster

AgentPhaseRole
situation-analyst1Read-only codebase explorer — maps files, interfaces, types, data flows
investigator2Deep-dive researcher — root causes, edge cases, integration points
architect3Software designer — approach, architecture, data model, test strategy
design-reviewer3bDesign quality gate — APPROVE or REVISE with findings
task-decomposer4Breaks design into numbered, dependency-aware tasks
task-reviewer4bTask list quality gate — APPROVE or REVISE
implementer5TDD developer — tests first, then code, one task at a time
impl-reviewer6Diff-based code reviewer — PASS, PASS_WITH_NOTES, or FAIL
comprehensive-reviewer7Cross-task holistic reviewer — naming, duplication, coherence
verifierFinalRuns full typecheck and test suite, fixes new failures

Invocation

Agents are invoked by the orchestrator using the Agent tool with the agent's name. The orchestrator passes only runtime parameters (workspace path, task number) — agent instructions are self-contained.

Phase Skipping by Effort Level

Which phases (and their agents) execute depends on the effort level, not task type:

EffortTemplateSkipped Phases
Slightphase-4b (task-reviewer), checkpoint-b, phase-7 (comprehensive-reviewer)
Mstandardphase-4b (task-reviewer), checkpoint-b
Lfull(none)

Model Configuration

Agents inherit the user's configured model by default — no model: key is set in agent frontmatter. This means agents run on whatever model the user has selected in their Claude Code configuration. To pin a specific agent to a particular model, add model: <name> to that agent's .md frontmatter.

Released under the MIT License.