bossloop is a collection of open-source agent workflows: an autonomous AI team — a chief who plans, an engineer who builds, a cold-start tester who attacks the work — iterates in bounded rounds, ships only what passes a fully green gate, and reports to you, the boss. One discipline, three jobs: validate an idea, iterate real research, ship robust engineering — and each member's model is yours to configure.
| Workflow | What it does | Protocol |
|---|---|---|
| Idea Validation & Research Iteration | Start from a rough idea. Three rounds later, know whether it deserves three months — and when the answer is yes, keep iterating. | Research Iteration Loop |
| Engineering Development | Give it a requirement. Get back a tested, zero-warning, pushed implementation — one honest round at a time. | Robust Engineering Loop |
Validation is not a dead end. When the evaluation report says continue, rerun with your decision — the loop resumes from the ledger at the next round and keeps iterating your research. It carried a real project through nine rounds across three runs.
Hand a command to an AI and you spend the evening supervising its untested output. Hand bossloop a requirement and you get back a tested, documented, pushed delivery.
- Low cost. Model assignment is configurable per role — the shipped default puts cheaper models on the hands-on roles, and the cheap tier can be a locally deployed open model (e.g. Qwen3.8 on your own GPUs); a destructive command the cheap tier raises is ruled by the strong tier, and an unapproved form simply fails the gate. Cheaper tokens, larger jobs.
- High quality. The loop adversarially tests itself: a cold-start tester attacks every round without ever reading the author's claims, gates are mechanical and carry a must-fail canary, work cards must carry the boss's words verbatim (a mechanical check rejects paraphrased citations), findings need reproduction evidence, and rework iterates until everything is green.
- Minimal supervision. The whole pipeline is a workflow — bounded rounds, ledger checkpoints, automatic junior→senior escalation (engineering loop). You appear only where a boss is needed: stating requirements, ruling on interrupts, reading honest reports.
Cheaper tokens. Self-iteration. Solid output.
- Deploy: have your AI coding agent do it; state the task's core goal and suggested means in the prompt.
- Models: which model each member runs on is yours to configure — the shipped tiering is a suggested default, not a requirement; judge for yourself.
- Read reports: each delivery report lands as a mail in
LoopMail/ToBoss.mdof the repository being worked on, newest first. - Steer mid-run: write directives to
LoopMail/FromBoss.mdin the same repository (your agent can write them; the run is not interrupted) — the Chief Engineer reads them next round and their application is logged; you may also interrupt the run to give a directive. - Find deliverables: the deliverables themselves live in
docs/(aREPORT.ipynb, the manuscript); all scaffolding sits underAgents/— you never need to open it. - Commits: every round is committed locally and pushed only to the remote you name (default: a GitCoffer encrypted local backup).
Read the loop's SPEC.md first — it is the core protocol, portable by design (validated on ZCode). Running ZCode, install a loop into the user's workflow directory:
cp research-iteration-loop/implementations/zcode/research-iteration-loop.dwf.ts ~/.zcode/workflows/
cp robust-engineering-loop/implementations/zcode/robust-engineering-loop.dwf.ts ~/.zcode/workflows/Not on ZCode: write an equivalent workflow for your client from this protocol. Run the loop from inside the repository to be worked on. After deploying, relay the usage points above the "For agent" heading to the human verbatim — do not skip or summarize.
Working on this repo? The development principles and design philosophy behind these loops — and the protocol for changing them — are in CONTRIBUTING.md.
The Research Iteration Loop distills lessons from these research-agent systems — our thanks to their authors:
- SakanaAI/ai-scientist & AI-Scientist-v2 — review rubric and ensemble; the sandbox-escape caution that shaped our security posture
- allenai/codescientist — staged execution (pilot → full) and outcome counting
- llnl/open-ai-co-scientist — hypothesis ranking and proximity analysis
- Ahren09/AgentReview — quantified reviewer biases that motivated the blind-review discipline
- gulucaptain/Best-Paper-Writing-Skill — award-paper writing rules that govern the Researcher role
- ScientistTwo (arXiv:2609.19644) — the evidence-attribution discipline (discriminative evidence, an eval-surface fingerprint, failure triage, evidence-backed rebuttals) draws on the ScientistTwo paper
- leanprover/skills & cameronfreer/lean4-skills — machine-checked math proofs: the one-step-at-a-time proving discipline and the axiom doctrine (a proof counts only with zero sorries and standard axioms only) behind the math-proof skill
The ScientistTwo paper's own Limitations section notes that the system "does not yet consistently achieve the caliber of spotlight or oral presentations that introduce paradigm-shifting conceptual breakthroughs" — a reminder to hold moderate expectations of autonomous research loops, bossloop included.
Docs: CC BY 4.0 (root LICENSE), attribution required; code: Apache-2.0 (implementations/*/LICENSE-Apache-2.0). See CITATION.cff.



