37 lines (29 loc) · 1.61 KB

Agent Backends

Configured via AGENT_BACKEND env var. Default: pydantic.

Backend	Architecture	Tokens (hello world MR)
`pydantic`	Parallel agents via asyncio.gather	~5K
`claudecode`	Single subprocess to `claude` CLI	~120-240K
`codex`	Single subprocess with JSON schema	~22K
`deepagents`	LLM orchestrator + subagents	~88K

Detailed docs:

Comparison (hello world MR, 2 files, 461 chars diff)

Backend	Model	Prompts	Findings	Tokens	Recommendation
pydantic	gpt-5.4	sec+logic+design (3)	0	5,297	approve
pydantic	gpt-4o-mini	sec+logic+design (3)	9	13,938	comment
codex	gpt-5.3-codex	common (1)	0	22,476	approve
deepagents	gpt-5.4	sec+logic+design (3)	0–1	88,528	approve/comment

When to use which

Scenario	Backend	Prompts	Why
CI pipeline (fast, cheap)	pydantic	common	parallel, low tokens
CI pipeline (deeper)	pydantic	security,logic,design	parallel, cheap
Claude Code (thorough)	claudecode	security,logic	file tools + git, structured output
Large MR (50+ files)	deepagents	security,logic,design	explores repo
Codex sandbox	codex	common	1 subprocess call

Adding a new agent backend: see adding_backends.md.