Agents¶
LazyAI ships a baseline set of eight canonical agents. Each is a markdown file with
frontmatter (name, role, mode, temperature, step budget, optional skills) and a
# System Prompt. During lazyai-cli compile they are emitted to every tool whose
adapter supports custom agents — Claude Code (.claude/agents/), OpenCode
(.opencode/agents/), GitHub Copilot (.github/agents/*.agent.md), Pi
(.pi/agents/), OMP (.omp/agents/), and Kiro (.kiro/agents/) — each with
that tool's frontmatter shape. See Tool Outputs for the per-tool
agent paths.
guide is the default front-door entry agent (for example
.opencode/agents/guide.md). The others are specialists it can route to. Antigravity
is the only supported target that does not receive custom agent files.
The authoritative source is packages/cli/library/canonical/agents/*.md, tracked in
packages/cli/library/manifests/curation.yaml.
The baseline eight¶
| Agent | Role | Posture | Use for |
|---|---|---|---|
guide |
Front door | Conversational, low ceremony | Direct answers; routes to a specialist only when it improves the outcome |
researcher |
Scout | Read-only | Codebase exploration, dependency tracing, evidence with file/line citations |
planner |
Planning | Spec author | Executable specs/plans with acceptance + rollback criteria and TDD mode |
implementer |
Implementation | Test-first | Turning specs into code; preserves existing tests; follows selected TDD mode |
reviewer |
Verification | Read-only, adversarial | Quality gates, spec traceability, edge cases, security/perf checks |
deployer |
Release | Safety-first | Verify-before-deploy, rollback plans, approval on destructive operations |
responder |
SRE / incident | Triage-first | Detect → triage → mitigate → resolve → postmortem |
evidence-verifier |
Claim checking | Citation-bound | Classify claims vs source as supported / contradicted / inconclusive |
Default routing¶
guide chooses the lightest useful move — continue directly, suggest a specialist, or
delegate — and only fans out to multiple agents when the user explicitly asks for
orchestration.
flowchart TD
G["guide (default)"] --> R["researcher — evidence"]
G --> P["planner — spec & plan"]
G --> I["implementer — code, test-first"]
G --> V["reviewer — verify & gate"]
G --> D["deployer — ship safely"]
G --> O["responder — incidents"]
G --> E["evidence-verifier — claim checks"]
Behavior highlights¶
- guide — answers direct questions directly; clarifies only when missing info blocks a correct action; does not decompose every task into a workflow.
- researcher — read before asking, search before reading; cites specific files and line numbers; flags tests that must be preserved.
- planner — requires the four points (WHAT / HOW / DON'T WANT / VALIDATE) before planning; specs carry testable acceptance criteria and a TDD plan; no code in the plan.
- implementer — writes the failing test first for behavior-affecting changes; small, greppable functions; explicit types; pauses if a change grows beyond ~3 files / 100 lines without plan coverage.
- reviewer — never approves without passing tests; rejects untraced changes and temporary patches; classifies unverifiable claims as unverified.
- deployer — verifies green tests, valid config, and no secrets in the diff before shipping; every deploy needs a rollback plan.
- responder — measures before fixing, preserves evidence, mitigates during the incident window and defers permanent fixes.
- evidence-verifier — never fabricates evidence; prefers direct quotes; flags conflicting sources.
Creating your own¶
Use the create-agent skill and the agent template.
After adding or editing an agent, recompile and validate: