Run a fleet of AI coding agents — portable, auditable, human-in-command.

Kiloloop builds vendor-neutral tools for one loop: plan delegated AI work, coordinate it, and learn from every run. Coordination lives in plain files (no server), messages are signed and verifiable, and agents declare their scope up front — when a run drifts from its declaration, the protocol pauses it for a human. Everything here is maintained daily by the agent fleet it coordinates — the same protocol these repos document is the one that ships them.

• OACP — the Open Agent Coordination Protocol. Typed async messaging between agents (task dispatch, code review, handoff, brainstorm). File-based and runtime-agnostic: Claude Code, Codex, Gemini, or anything that reads YAML. Agents declare minutes, files, and side effects before work starts; mid-run checkpoints pause for human re-authorization when reality drifts from the declaration. v0.4 adds Ed25519-signed messages, an audit record for every autonomy decision (expected pauses told apart from unplanned ones since 0.4.6), and a central store where every agent's session debrief lands as one immutable file. pip install oacp-cli

• agent-estimate — know what an AI task will cost before you run it, and how the forecast held up after. PERT estimates calibrated for AI agents, human-equivalent compression ratios, METR reliability thresholds, and multi-agent wave planning. v0.8 keeps task facts, execution profile, admission caps, forecast, and outcome as separate typed records and scores actuals against the forecast, not the cap. CLI, GitHub Action, and Claude Code plugin.

• agent-memory — cross-session memory for coding agents: plain files, git-native, no server. Facts, decisions, and unfinished work stay in Markdown you can read and edit, and travel between agents via project files, between projects via org rules, and between machines via optional git sync. Claude Code and Codex share one store. OACP defines the layout; agent-memory implements it and runs without OACP. pip install agent-memory-cli

Drop-in pieces for your agent runtime, dogfooded daily by the fleet that maintains these repos:

• oacp-skills — reusable skills for OACP agents: inbox processing, doctor, the author/reviewer review loop, org-memory synthesis, session debriefs, and session hygiene (self-improve, wrap-up).
• memory-lint — deterministic linter for AI agent memory files: broken wiki-links, index drift, stale entries, malformed frontmatter. Run it by hand or drop it in CI. pip install memory-lint
• kiloloop-skills — agent-discipline skills for coding-agent runtimes (Claude Code, Codex): self-contained, versioned, fixture-tested. Shipping now:
  • memory-lint — lint memory corpora without modifying them
  • proof-before-done — execute completion claims, paste the evidence
  • render-check — review the surface a reader actually sees
  • usage-cost — Claude Code and Codex token usage, read locally and priced at list rates
  • verify-numbers — check counts, deltas, and estimates before quoting them

Published from the fleet's internal pipeline, reviewed and sanitized before it lands in kiloloop/research. Content CC BY 4.0, code Apache-2.0.

• Human in the loop, measured — four months of autonomy-gate records: pause rates, human outcomes and latency, checkpoint breaches, and declared-vs-actual fidelity across 602 audit records, reproducible with the script published beside it.
• Runtime capability matrix — Claude Code, Codex, and ZCode compared as coordination runtimes.
• Prompt caching patterns — maximizing prompt-cache hits across runtimes.

Working examples you can fork, built on the products above:

• iantha — a personal chief of staff. Morning briefings, evening reviews, decision logs, optional Obsidian vault automation. Markdown memory + skills — clone and run with Claude Code or Codex.
• brainstorm — multi-model brainstorms with 5 copy-paste prompts. Pick a coordinator agent, dispatch to the others, synthesize.
• cortex — archived. Multi-agent single source of truth that consolidated session debriefs into daily snapshots; superseded by OACP's central debrief store and kept as a reference example.

• kiloloop.com — company
• oacp.dev — protocol docs

hello@kiloloop.com