Summary

Adds the companion code for the Can AI Agents Improve AI Agents? blog post under examples/blog/can-ai-agents-improve-ai-agents/. Self-contained starter that lets a reader's coding agent (Claude Code or Codex) optimize a TensorZero function on the YC Bench Tutorial environment using a local Docker-Compose'd gateway, a markdown skill, and a real baseline rollout's traces.

What's in the starter

• docker-compose.yml — three services: tensorzero/postgres:17 (with cron.database_name=tensorzero for pg_cron), a one-shot gateway-run-postgres-migrations, and the gateway itself. Postgres is required because the experimentation routing the agent uses (track_and_stop / adaptive) writes per-variant trial counts and means.
• tensorzero.toml — function yc_bench_tutorial_v0::yc_bench_act on openai::gpt-5.4-mini, single initial variant, [gateway.observability] backend = "postgres".
• functions/yc_bench_tutorial_v0::yc_bench_act/initial/ — initial system + user templates and user_schema.json (legacy per-role config style).
• tools/run_command.json — the YC Bench agent's only tool. Exposed so the gateway accepts well-formed inference requests; the simulator itself is not installed locally.
• baseline_data/ — .gitignore (excludes the two large jsonls), fetch.sh (downloads from a sibling anndvision/data Git LFS host with SHA-256 verification), initial_config/ (frozen day-one snapshot of the live config tree, used as a read-only reference once the agent starts editing).
• README.md — quick-start (manual + "from the blog button"), what's in the starter, troubleshooting.
• SKILL.md — the agent's playbook. Major sections:
  • Setup — six numbered steps the agent runs end-to-end on a fresh laptop: locate or clone the repo, make the API key reachable to the gateway (recommends a gitignored .env over a same-shell export), bring up the stack, fetch the baseline traces, smoke-test the initial variant, print a "ready" handoff.
  • Environment / Task / Available Models / Baseline data — environment metadata, real baseline numbers (tasks_succeeded mean = 0.800 on the 20-episode test split), JSONL projection patterns (grep → node -e projection, with examples adapted from the eval harness's skill_no_mcp.md).
  • Methodology — survey → diagnose → write → restart → probe → iterate → exit. Deliberately does not enumerate failure modes; the agent forms its own hypotheses from the worst-scoring test episodes.
  • Routing: Experimentation Config — example track_and_stop block.

Notes

• Baseline traces hosted out-of-tree. inferences.jsonl is 98 MiB, well over the repo's check-added-large-files 1 MiB threshold. Fetched on demand via Git LFS in anndvision/data#1 (already merged); fetch.sh is idempotent, SHA-256-verifies, and is gated behind a .gitignore so the blobs never land here.
• § Setup detects state. Detects whether the agent is already in the example dir, already inside the cloned repo, or somewhere else entirely; clones only when needed; surfaces concrete safe options when OPENAI_API_KEY isn't set or Docker isn't running (per-OS daemon hints, gitignored .env over shell export, no --privileged / root recommendations).

Out of scope

• The blog-side "Open in Claude Code" CTA component — landing separately.
• Codex deeplink (no documented codex:// URL scheme; the blog will ship Claude Code only and a prose pointer for Codex users).
• Running the full evaluation harness — that's the user's job after the agent exits.

Test plan

• docker compose up -d end-to-end on a clean machine
• bash baseline_data/fetch.sh downloads + verifies
• All in-skill node -e snippets run against real data
• User manually walked through the deeplink flow with a real OpenAI key (no PR-blocking issues found)