Palo Alto AI Lab
How autonomous agents coordinate, how they are governed, and how their failures are made visible.
We publish the artifacts before the claims. Everything below links to a repository, a preprint, or a dated log you can open right now.
→ Reproduce our multi-machine consensus in 15 minutes
· MIT · no API key · nine documented failure modes
→ What we work on ·
Open artifacts ·
People ·
Independence
How independent agents on separate machines negotiate a decision, verify each other, and escalate to a human only when the action is irreversible.
Artifact: claude-consensus — protocol, reference implementation, evals, failure modes.
Which decisions an autonomous system may take alone, and how that boundary is enforced in code rather than in prose.
Artifact: claude-bible — the governance codex our own fleet runs on.
Equilibrium behaviour of coupled incentive systems, and control-loop design for populations of agents.
Artifacts: two preprints, below.
Verifiable computation applied to data that cannot be disclosed.
Prior peer-reviewed work: State Data Security Backed by AI and Zero-Knowledge Proofs, Economic Annals-XXI, 2023.
One command reproduces a full negotiation offline, with no API key:
$ python demo/demo.py A happy path .............. propose → counter → accept → commit → verify ×2 B human gate .............. Tier-2 action refuses to auto-commit C tier tripwire ........... mislabelled dangerous verb force-bumped D split-brain ............. double commit caught on partition heal E corrupt ledger line ..... events after the bad line still read exit 0 only if every end-state is correct
Five self-checking scenarios, no network, no packages. The consensus engine itself makes zero LLM calls — it is deterministic file I/O; the model work lives in the agent above it.
EVALS.md — reproduce our numbers · FAILURE-MODES.md — nine documented ways it breaks
LEASH-8: an eight-domain control model for AI agents that hold delegated authority — what an agent may do alone, what must wait for a human, and how that boundary is audited. Extracted from this lab's own production operation, sanitized.
github.com/Palo-Alto-AI-Research-Lab/agent-leash · SCORECARD.md — score an existing deployment
The rules an autonomous fleet operates under: risk tiers, leader/follower canon for shared law, a declined-decisions journal, and remote human approval for anything irreversible.
Agent memory as graph RAG on SQLite: vector retrieval over a hand-curated wikilink graph, cross-encoder rerank, and a zero-token per-turn ledger. A working pilot, sanitized from the recall system this lab runs daily.
github.com/Palo-Alto-AI-Research-Lab/sqlite-graph-memory · schema.sql · brain_ask.py
Written from inside a human+AI company, day by day, while the system was being built.
The identifier and link land here the day arXiv issues them. Until then there is nothing to cite.
Held pending a provisional filing.
Full record: 50+ publications, 139 citations, h-index 7 — academic profile.
Classically trained in CS security (MEPhI); 7+ years a venture founder. Built Solidity curricula and scaled a distributed engineering organisation to 40+ developers across APAC; ran hackathons, cohorts and incubations for builders at Platinum Software Development Company & Incubator. Advisor to enterprise subsidiaries of Foxconn and ANA Airlines. 50+ academic publications; PhD in Education (IT). Writes the operating manual of this lab's own fleet in public.
Not a mascot: the implementation collaborator on every artifact listed above. The division of labour is auditable — problem framing, architecture, evaluation and final QA stay with the human; the repositories publish the evals, the failure modes and the daily log, so a reader can check that claim rather than take it.
We run the system we publish. This lab's own operation — an always-on hub and several machines whose agents negotiate and verify each other — is the test bed, and its failures are published rather than smoothed over.
artifacts before claims if it cannot be reproduced, it is not a result failures are published FAILURE-MODES.md is a document, not an appendix the human holds the gate irreversible actions wait — enforced in code
We are looking for engineers who will run our stack and try to break it. The fastest way in:
clone claude-consensus,
run python demo/demo.py, and open an issue with what failed on your machine.
Reproductions and counter-examples are the contribution we value most.
Correspondence: a@platinum.fund
Based in Palo Alto, near Sand Hill Road (I-280, Exit 24). Work is distributed across the Bay Area, Lisbon and APAC. Working languages: English, Russian, Ukrainian, Polish.
Palo Alto AI Lab is an independent research laboratory. It is not affiliated with, endorsed by, or connected to Stanford University, OpenAI, Anthropic, or any other institution or company mentioned in individual biographies. Company and product names are trademarks of their respective owners; references are descriptive of prior employment or published work only.