Agents that write agents.

Give tamago one line about a job and it hatches a complete agent: a role, a production system prompt, a tool allowlist inside a declared risk budget, an escalation rule, and an eval suite it must pass. Then it can generate that agent for real runtimes, run it, score it, and grow it generation by generation.

tamago means egg (卵). Agents hatch agents.

tamago new "summarize merged PRs into a weekly changelog"

An LLM designer (Claude Opus 4.8) turns the job into an AgentSpec: name, role, system prompt, tools, risk class, model tier, triggers, and 3 to 5 eval fixtures born with the agent. A risk gate refuses any design whose tools exceed its declared class, so a read-only agent can never come out holding git_push.

The spec lands in the garden at ~/.tamago/garden/NAME/ with its full generation history.

tamago eval changelog-scribe     # run the fixtures, an LLM judge scores 0-100 each
tamago gen changelog-scribe      # emit runnable artifacts
tamago run changelog-scribe "..."
tamago grow changelog-scribe     # eval, rewrite, keep the better generation

grow reruns the suite, feeds the failing transcripts to a rewriter, saves the candidate as generation N+1, re-evals, and keeps it only when the mean score actually improves.

Bare tamago opens the dashboard: the garden on the left, the selected agent on the right, live streaming output for hatch, run, eval, and grow.

Keys: j/k move, enter spec, n new, r run, e eval, g grow, esc cancel, q quit.

Generation is deterministic: no LLM at gen time, the same spec always produces the same files.

name: changelog-scribe
job: summarize merged PRs into a weekly changelog
role: a careful release-notes writer ...
risk: read          # read | write | admin, gates the tool allowlist
tier: standard      # fast | standard | deep -> haiku | sonnet | opus
system_prompt: |
  ...
tools: [web_fetch, read_file]
triggers:
  - type: cron
    cron: "0 9 * * MON"
escalation: when a PR looks like a security fix
evals:
  - input: three PRs merged, one breaking
    expect: groups by area and flags the break prominently
meta:
  generation: 3
  parent: release-notes-bot   # lineage, from `tamago new --from`

go install github.com/tamnd/tamago@latest
export ANTHROPIC_API_KEY=...

The designer, judge, and rewriter all run on the real API from the first command; there is no offline or mock mode.

tamago speaks two wires: the Anthropic API and any OpenAI-compatible chat completions server.

Selection order: TAMAGO_PROVIDER (anthropic or openai) wins, else ANTHROPIC_API_KEY picks Anthropic, else OPENAI_API_KEY picks OpenAI-compatible, else Anthropic.

export TAMAGO_PROVIDER=openai
export OPENAI_BASE_URL=http://localhost:8080/v1   # default https://api.openai.com/v1
export OPENAI_API_KEY=...

Many OpenAI-compatible servers ignore response_format, so tamago never relies on it: JSON answers are requested by embedding the schema in the prompt, extracted tolerantly from chatty replies, and retried once with the parse error when the first attempt is off. Transient 429 and 5xx responses are retried with a short backoff before any output has streamed.

MIT