AI Agents

Trellis treats the agent's experience graph—artifacts, tool outputs, rewards, sibling comparisons, causal lineage—as first-class, governed, queryable database state rather than disposable JSON checkpoints. The framing is sharp: frontier selection becomes a query, cross-session reuse becomes vector-seeded graph retrieval, training-data extraction becomes a materialized view, and reconstructing past agent state becomes a time-travel query. When the database owns the graph, agents become stateless compute and crash recovery plus horizontal scaling fall out for free. Grounded in KernelEvolve, a production kernel optimizer at Meta.

Memory consolidation manufactures confidence: a casual, hedged remark gets rewritten into a flat dated assertion that later steps obey like verified fact—no attacker needed. The agent responds to phrasing confidence, not source: attributed, unattributed, and forged "system of record" claims all grant alike, while the evidential register ("reportedly") is obeyed like a flat assertion on most models. Passive "unverified" tags are ignored; active "do not trust this" escalates even correct memory. The real fix is keeping tentative phrasing in the store, plus one redundant source.

TraceLab releases roughly 4,300 real coding-agent sessions—about 350K LLM steps and 430K tool calls—collected from day-to-day Claude Code and Codex use, filling a genuine gap since public traces don't capture serving-relevant workload patterns. The analysis finds long autonomous loops, long contexts with short outputs, heavily-tailed tool-call distributions, and high-but-imperfect prefix cache hit rates. These point to concrete serving wins: lower-overhead tool calling, append-length-aware prefill, semantic tool-latency prediction, and KV-cache management around human-paced gaps. Dataset, pipeline, and analysis code are all released.

Qwen-AgentWorld trains language world models that simulate agentic environments across seven domains via long chain-of-thought, in 35B-A3B and 397B-A17B MoE sizes, from 10M+ real interaction trajectories. The three-stage pipeline—CPT for state-transition dynamics, SFT for next-state-prediction reasoning, RL with hybrid rubric-and-rule rewards for fidelity—is paired with AgentWorldBench, built from five frontier models on nine benchmarks. The decoupled-simulator use case is the interesting part: a learned environment model for scaling rollouts without live execution. Self-reported wins over frontier models warrant independent checking.

Agents-A1 is a 35B MoE that claims trillion-parameter-level results by scaling agent horizon rather than parameters, training on trajectories averaging 45K tokens. The recipe: full-domain SFT, per-domain teacher models, then multi-teacher domain-routed on-policy distillation with salient vocabulary alignment to unify six domains into one deployable student. Reported numbers beat or match Kimi-K2.6 and DeepSeek-V4-pro on SEAL-0 (56.4), IFBench (80.6), and FrontierScience-Olympiad (79.0). The distillation method is the real contribution; treat the unfamiliar competitor model names with caution.

Husain reframes the common "our product is hard to eval" objection as a product defect: if you can't verify an output cheaply, users have to redo the work from scratch to trust it. Drawing on three products he advised—including an internal AI data agent answering business questions—he shows before/after sketches arguing that designing for verifiability comes before building evals. The point lands because it ties eval difficulty to UX surface area, not measurement tooling.

DeepReinforce's first release is an MIT-licensed coding model built on Gemma 4 and Qwen 3.5 (both Apache 2.0), shipping in 9B/31B dense and 35B/397B MoE variants with self-scaffolding for agentic coding. Willison ran the 35B Q4 GGUF (20GB) in LM Studio against a Datasette checkout, finding it competent over many tool calls—locating cookie-decoding code and dialog handlers on request. Promising open weights, but these are hands-on impressions, not measured agent-harness benchmarks.