Open-source systems for the agent era.
The trust, memory, and verification layer autonomous agents are missing. Twelve systems — agents now act, but almost nothing verifies what they did, proves a human was behind it, or keeps their world-state coherent. Each is its own probe into that frontier.
Git-diff for agent behavior. Swap a model, prompt, or tool and see the exact decision fork that changed — as a reviewable artifact in your PR, not a customer complaint three weeks later.
The Sigstore for agent execution. Turn raw production agent logs into independently verifiable, privacy-scrubbed, attestable records — the missing unit of account for what an agent actually did.
Proof-of-human for the agent era. Model the generative physics of human motor control, not its surface statistics — so an AI can't pass by mimicking the numbers. Born in anti-cheat, sells into fraud, RPA, and CAPTCHA-replacement.
Long-horizon agents silently rot: a fact gets revised at hour 47 and a dozen downstream decisions built on the old one never get flagged. foghorn makes world-state versioned, mergeable, and staleness-aware.
When many agents mutate one shared world, last-writer-wins quietly destroys state. agentcrdt makes the conflict a first-class event — so contradictions surface (or become emergent story) instead of vanishing.
A long-running agent's beliefs drift out of sync with the world it lives in — it talks about a king the player watched die. worldoracle reconciles agent memory against ground truth before it speaks.
Every robot builds a private model of a space and throws it away on reboot; the next robot re-maps from zero. polaroid is a linkable, server-less spatial memory many agents can read and write at once.
Computer-use agents are graded by screenshot-judging vision models that pass an agent reaching a visually-correct state via a wrong, dangerous path. groundcrew reads the OS for deterministic before/after action receipts.
Every GUI-agent skill is a frozen snapshot of how an app behaved on authoring day. SaaS UIs ship weekly, so skills silently rot. clickproof is the freshness oracle that tells an agent whether a UI fact is still true today.
The thing AI Dungeon failed at: a game master that is rules-faithful over many hours. rulegraph parses any rulebook into a constraint engine and only routes the genuinely ambiguous calls to the model.
Economy exploits — arbitrage loops, dupe-adjacent flows, bot-amplifiable farms — surface after launch, in the wild. balancelab finds them first with adversarial agents reasoning over your rule graph.
Multi-agent NPC societies need norms that actually bind — without hand-scripting every rule. normsync lets LLMs decide what norms exist and a deterministic engine decide what they enforce.