Write Claude Code subagents once. Emit native agent outputs everywhere.
CaesarAgent is an AI agent harness for Claude Code subagents, AGENTS.md-compatible tools, and native agent outputs across the modern AI coding stack.
Every AI coding tool invented its own agent file format. Maintaining the same backend reviewer, Rust expert, or security auditor prompt across multiple tools by hand creates drift and stale behavior.
CaesarAgent keeps one canonical Markdown + YAML definition per expert agent, then transpiles it into each tool's native format with schema validation on every build. One source of truth. No per-tool copy-paste. No prompt drift.
pnpm install
pnpm build
pnpm validateAfter building the CLI, call the built entrypoint and install one category into a target project:
node packages/cli/dist/index.js build
node packages/cli/dist/index.js install 05-data-ai --tool claude --dest <target-project> # → .claude/agents/*.md
node packages/cli/dist/index.js install 03-infrastructure --tool codex --dest <target-project>
node packages/cli/dist/index.js validate --strictReplace <target-project> with the project directory that should receive the generated agent files. Category accepts NN, NN-name, or name (for example: 05, 05-data-ai, data-ai).
| Tool | Output | Tier |
|---|---|---|
| Claude Code | .claude/agents/*.md |
native |
| OpenCode | .opencode/agents/*.md |
native |
| Kiro | .kiro/agents/*.json |
native |
| Codex | .codex/agents/*.toml |
native |
| Factory / Droid | .factory/droids/*.md |
native |
| GitHub Copilot / VS Code | .github/agents/*.agent.md |
native |
| Gemini CLI | .gemini/agents/*.md |
native, inherit-only emit |
| OpenHands | .agents/skills/*/SKILL.md |
native skill output |
| Kilo Code | .kilocodemodes |
native aggregate output |
| Cursor, Windsurf, Cline, Antigravity, Amp | shared AGENTS.md routing index |
fallback |
| Roo Code | not emitted | skipped |
Full matrix and per-tool limits: docs/tool-support-matrix.md.
CaesarAgent currently ships 134 expert coding agents across 10 categories.
| # | Category | Count | Examples |
|---|---|---|---|
| 01 | core-development | 7 | backend-developer, api-designer, microservices-architect |
| 02 | language-specialists | 28 | typescript-pro, rust-engineer, python-pro, flutter-expert |
| 03 | infrastructure | 18 | kubernetes-specialist, terraform-engineer, sre-engineer |
| 04 | quality-security | 15 | code-reviewer, security-auditor, penetration-tester |
| 05 | data-ai | 19 | ai-engineer, ml-engineer, llm-architect, eval-engineer |
| 06 | developer-experience | 12 | refactoring-specialist, mcp-developer, build-engineer |
| 07 | specialized-domains | 11 | blockchain-developer, fintech-engineer, embedded-systems |
| 08 | business-product | 8 | product-manager, business-analyst, technical-writer |
| 09 | meta-orchestration | 10 | multi-agent-coordinator, context-manager, policy-guardrail-designer |
| 10 | research-analysis | 6 | research-analyst, search-specialist, first-principles-thinking |
Each category can ship as its own npm package of prebuilt artifacts (@caesar/lang, @caesar/infra, and others) so projects install only the roles they need.
agents/{NN-category}/{agent}.md canonical source (Markdown + YAML)
│
▼ @caesar/agents-core (schema → transpiler → emitters)
│
▼ per-tool output validators
│
▼
dist/{tool}/… native files for each tool + dist/agents-md/AGENTS.md
Architecture detail: docs/system-architecture.md.
| Package | Purpose |
|---|---|
@caesar/agents-core |
Canonical schema, transpiler engine, emitters, and output validators |
@caesar/cli |
The caesar CLI: build, validate, and install |
@caesar/{category} |
Prebuilt agent artifacts per category, generated from dist/ at release |
For tools that read repository-level instructions instead of tool-native agent files, CaesarAgent emits a shared AGENTS.md routing index. That fallback keeps role discovery consistent, but native targets remain the preferred path when a tool supports isolated agents, permissions, or model routing.
Every agent goes through a 6-gate quality pipeline (see docs/g1-research-playbook.md and docs/system-architecture.md):
- G1 Research — current primary sources plus corroborating engineering sources.
- G2 Draft — copy
agents/_template/canonical-agent-template.md; fill the frontmatter superset and all 6 body sections. - G3 Self-review — tool minimalism, model tier, crisp routing, no sibling-trigger overlap.
- G4 Expert review — measurable rubric, read-only poka-yoke for auditors.
- G5 Transpile + validate —
caesar validate --strictmust pass. - G6 Smoke test — emitted files load and parse for every target.
Local validation:
pnpm build
pnpm test
pnpm validate
pnpm lintThey are specialized agent definitions that Claude Code can route to for focused work such as code review, infrastructure, data engineering, or security analysis.
Not directly as the same file. CaesarAgent solves that by compiling one canonical source into native OpenCode, Codex, and other target formats.
AGENTS.md is the shared repository instruction index for compatible tools and fallback routing. It is useful for consistency, but it does not replace native tool support where native agents are available.
The canonical agent is the only source that maintainers edit. Generated outputs are rebuilt and validated instead of hand-maintained per tool.
No. It is a harness, transpiler, and curated software-engineering agent library for maintainers who want repeatable multi-tool outputs.
GPL-3.0-only. See LICENSE.