Show HN: AgentRun – open-source, Self-improving Managed Agents
TL;DR Highlight
Agents save notes and skills after each run, cutting code review cost to 18 cents
Who Should Read
A backend developer worried about token costs and crash recovery for agents that repeat the same jobs
Core Mechanics
- Agents keep what they learn
- With "learning": true, the agent proposes notes, skills or workflows on its last turn of each run.
- Code checks them and a grader scores them, and only the keepers are committed to the agent's git folder on main.
- The next run starts from there.
- Code review cost 18 cents
- Greptile started charging $1 per review in March, and their monthly bill went from $180 to $6,451.
- They built their own reviewer on a cheap model and tuned it on their own PRs, which costs 18 cents a review.
- On 30 unseen PRs it caught 1.6x the real bugs Greptile did, they say.
- Opus taught a cheap model
- Opus worked out how to classify merchants, then handed what it learned to a cheap model.
- The same job dropped from $0.89 to $0.12 per merchant.
- Not everything needs an agent
- Write repeatable work as a JSON workflow of tool calls, model steps and typed fast decisions with confidence.
- Only unclear cases fall back to an agent.
- After a learned skill serves 3 runs, the run drafts a workflow, which is dry-run with no model or tool first.
- Runs survive crashes
- Every run lives in one store (SQLite or Postgres), so a restarted server carries on.
- Finished steps never rerun, and a write cut off mid-call is marked interrupted instead of retried.
- Controls come built in
- Each tool is allowed, asks first, or is denied, and each run can have a spend cap.
- At the cap a run pauses instead of failing, and unpriced costs show as unknown, never $0.
- Harnesses supported: pi, Claude Code, Codex, OpenCode.
Community Reactions
- Question about subscriptions
- One commenter asked whether it uses the Claude Code harness or bypasses it, because of subscriptions.
- Request to explain self-improvement
- A commenter asked how the self-improving part works.
- Local orchestrator suggested instead
- One commenter said that if your team uses subscriptions, a local orchestrator for PR reviews is cheaper and higher quality.
- They pointed to their own project, codealong.
- Similar agent built independently
- A commenter built a similar Python agent that finds other agents at runtime on Modal and has a safety layer checking every action.
- They also liked the Mission Control screen.
How to Apply
- Run the KYB template today
- With Node 24.22.3+, run npx @parcha/agentrun and add OPENROUTER_API_KEY in Settings.
- Apply the KYB due diligence template and start a run with a company name and website.
- When it finishes, open the Learning tab to see what the agent kept.
- Turn on learning for a repeated agent
- Add "learning": true to the definition of an agent that does the same job dozens of times.
- After a few runs, check notes and skills in the Learning tab and retire unneeded ones.
- Put a spend cap on runs
- Add --max-cost-usd 1 to runs start.
- When a run pauses at the cap, read the minimum it needs and decide whether to continue.
- Move fixed steps into a workflow
- Write the always-same-order part of an agent job as a JSON workflow.
- Validate it with npx @parcha/agentrun workflows validate and route only unclear cases to the agent.
Try It
# Paste into a terminal (second terminal while the server is running, Node 24.22.3+)
cat > agent.json <<'EOF'
{"definition": {"id": "notes", "version": 1, "name": "Notes", "instructions": "Do what you are asked, and write what you make to a file.", "skills": [], "tools": ["write", "ask_person"]}}
EOF
npx @parcha/agentrun agents create --body @agent.json
npx @parcha/agentrun runs start notes --max-cost-usd 1 --wait --body '{"input": {"content": "Ask me which region to cover, then write region.md about it."}}' > out.json
run=$(node -p 'require("./out.json").run_id')
npx @parcha/agentrun runs message "$run" --content "Europe"
npx @parcha/agentrun runs wait "$run" > out.json
npx @parcha/agentrun runs read-file "$run" region.md
# Next: add "learning": true to the agent definition, run it several times, and check the Learning tabTerminology
Related Papers
LLM assisted reverse engineering
Pointing an agent at binaries with skills and MD notes made years-long hacks feasible
Metrics-in-the-Loop: Complexity Feedback Reshapes Agent-Written Code
Injecting per-turn complexity reports into coding agents cut code erosion without changing correctness
Why is Gemini 4 Carbon taking so long?
Carbon took nearly an hour on bug fixes that Astra and Opus finished in minutes, mostly from extra unrequested work.
How the Claude Code team thinks about verification
E2E, property and differential tests plus verification rules let Claude run in parallel.
Running coding agents without handing them my credentials
Credentials stay on the laptop; agents run commands there via Touch ID approval
Windows 11 Is Collecting Data on Which Coding Agents and Developer Tools You Use
Reverse engineering shows Windows 11 logs which coding agents you use and reports it as optional diagnostics
The Silent Cost of Conversation History in LLM Reasoning
Keeping full conversation history dropped accuracy from 0.6 to 0.1 within 20 turns; a verified state summary prevented it.
Anthropic upgrades Claude Code Projects to run parallel AI workers
Claude Code Projects now has a coordinator that splits work into parallel threads and opens PRs.