Shipped agentic systems
that run in production.
The trust, memory, and verification layer that autonomous agents are missing. Twelve open-source systems — verify what an agent did, prove a human was behind it, and keep agent world-state coherent. Built to ship, and built to run in the real world.
SYSTEM STATE
AGENTIC OPS
OBSERVABILITY
AGENT NETWORK
RUNTIME HEALTH
Current Focus
agentdelta — behavioral version control for agents
Git-diff for agent behavior. Swap a model, prompt, or tool and see the exact decision fork that changed — a semantic regression diff posted on the pull request, before a customer finds it.
humanproof — proof-of-human from the physics of motor control
Detect AI input by modeling the human neuromuscular plant, not its surface statistics — so an AI can't pass by matching a histogram. Born in anti-cheat; generalizes to fraud, RPA, and CAPTCHA-replacement.
worldgit — git for agent world-state
Long-horizon agents silently rot: a fact is revised and a dozen downstream decisions built on the old one are never flagged. worldgit version-controls world-state and propagates staleness — Make/Bazel semantics for agent cognition.
The trust layer agents are missing
Prove what an agent did — to a reviewer, an auditor, a regulator.
agentdelta turns two stochastic runs into a reviewable semantic regression diff in your pull request. tracemarket turns raw production logs into canonicalized, privacy-scrubbed, hash-chained, independently verifiable records — the Sigstore for agent execution.
Was a human behind this action?
humanproof models the physics of human motor control, not its surface statistics — so an AI can't pass by matching a histogram. Born in anti-cheat; generalizes to fraud, RPA, and CAPTCHA-replacement.
Keep agents' beliefs coherent over the long horizon.
worldgit version-controls world-state and propagates staleness so a revised fact flags every decision built on the old one; agentcrdt turns concurrent conflicts into observable events instead of silent overwrites.
The Labs
Verification & Trust
Behavioral diffs and verifiable, attestable execution traces — proving what an agent actually did.
Proof-of-Human
Telling human from machine by the physics of motor control — anti-cheat, fraud, and identity for the agent era.
Memory & World-State
Version-controlled, conflict-aware, contradiction-detecting memory so agents and worlds stay coherent.
Computer-Use
Ground-truth state oracles and version-pinned UI facts for agents that act on real software.
Game Intelligence
Rules-faithful game masters, economy red-teaming, and emergent norm enforcement for agent-driven games.
How it all connects
Latest Briefings
The agent coordination problem
May 20Proof-of-human for the agent era
May 10Observability for autonomous systems
Apr 28Infrastructure that thinks
Apr 15Enterthesystem.
Operational systems designed for real-world decision making at scale.