Skip to content
howdoaiagentswork.com

Agent Signal Log

A dated, sourced log of what happened in AI agents — and the mechanism behind each event. Entries link to the original source, never a re-upload.

  13. newslaunch

    Laya: the open-source answer to Jev hit #1 on Hugging Face in 48 hours

    Jev's Google Trends reading fell from +59,400% to +3,750% in two days, but the topic did not leave - it moved to its open-source clone. Laya answers typed questions (choice, score, noul) over any state in one non-autoregressive forward pass and returns probabilities instead of text, so there is no string-generation surface to hallucinate on. Three checkpoints (421M English, 322M multilingual over 100+ languages, one named laya-typed-decisions), Apache-2.0, ~33 ms on a T4. The mechanism worth recording is not the speed: it is that a closed API became a weekend replication target. Hugging Face's CEO called it the #1 trending model on the platform; the benchmark numbers saying it beats Jev are the author's own, and AINews had already flagged the counter-claims - RLCD asserted without justification, and confidence that is entropy-based rather than calibrated.

    Source: The Left Shift - Laya claims the top spot on Hugging Face · Latent Space / AINews - Here are 6 clones of Jev in 2 days

    Deep dive →

  15. newsframework

    Google AX makes the agent, not the container, the unit of isolation

    AX is Google's declarative runtime for agent workloads, and its four primitives are the argument: Task (sandboxed execution with CPU and memory limits), Workspace (pre-wired git repos, MCP servers and skill packs so each run starts warm), Gateway (an outbound allowlist that defaults to deny and injects credentials at the edge), and Model (provider plus secrets). Applied with kubectl-shaped commands: ax apply, ax get, ax watch, ax suspend, ax resume, ax ssh. The v0.3.0 rewrite moved task state out of Kubernetes etcd into Redis Streams, because etcd was never built for millions of short-lived agent tasks. Suspend checkpoints a waiting agent and resume restarts it from the interruption - the cost control that matters when an agent idles on a model API. The caveats are on the record too: the README still warns of breaking changes before a stable release, early testers report dropped egress-proxy connections and thin secrets management, and the billions-of-workloads-per-cluster claim ships without a reproducible benchmark.

    Source: AX - official docs · AI Insiders - Google quietly open sources a Kubernetes-style orchestrator for AI agents

    Deep dive →

  20. newslaunch

    TypeSafe Jev: a model that cannot hallucinate because it cannot write

    Jev skips string generation entirely — unstructured state in, typed probabilistic decisions out, 70–500ms, output tokens free. No text means no hallucination surface. A week after launch it went GA; Vercel's own paid-team stats put AI Gateway coverage at 27.8%. TypeSafe's workflow scores are self-reported, with no independent audit; the founder says they can't prove pricing isn't subsidized; Vercel tells developers to calibrate the returned probabilities on their own labeled samples first. TypeSafe's listed weak spots are counting, numeric ops, and date comparison — leave those to code.

    Source: TypeSafe AI — Introducing System One Models and Jev · Latent Space — Here are 6 Clones of Jev in 2 days