Quick Hits

• A decision-model race is forming: @omarsar0 flags CLM (Contrastive Language Model) as the second "System One" model after Jev, with @jackyk02's announcement claiming 9x faster inference and new marks on agentic coding benchmarks. The numbers are self-reported, so treat them as claims until outside evals appear.
• @muratcan says recursive self-improvement is visibly working in his team's harness on Opus 5.5 and GPT-6 Astra: agents write evals, call each other, grade, and review, while humans still own merge and deploy "for now."
• Docker and the Linux Foundation announced the Docker Sandbox Kit Specification at WeAreDevs: an Apache 2.0 open standard for declaring what an agent may do, where it may reach, and what it may touch.
• Per @kimmonismus, Trump wants to "leave it exactly where it is" on AI, says China agrees, and calls the DOJ the guardrail, which @kimmonismus reads as the end of the slowdown discussion.
• @suraj_sharma14 surfaces Meta's Global AI Developer Hackathon: 10 days, fully virtual, free, open worldwide, with $1M in prizes and access to Meta's newest models.

The "System One" decision layer is suddenly contested

The fast-decision slot inside custom harnesses went from one contender to two this week, and practitioners are building routing stacks around it.

@omarsar0's framing: watch this wave if you build custom harnesses. CLM is 9x faster than Jev and, in his read, a better verifier on long-horizon tasks. The two differ in method: Jev is trained with Reinforcement Learning for Calibrated Decisions and returns typed decisions with probabilities from predefined questions, while CLM embeds situations and candidate actions, then ranks or selects the best match. Per @jackyk02's thread, CLM-8B is pre-trained on internet-scale data, matches Jev on computer-use, gaming, and tool-calling, and sets claimed SOTA on DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%), benchmarks where Jev "fails to serve as an effective verifier." The team also disaggregates state and action embeddings for caching, cutting latency when state evolves but the action set is fixed.

The practical payoff shows up in @Av1dlive's workflow: pairing Jev with Opus 5.5 cut his costs and time by roughly 80%. Jev picks from options the harness prepares (which notes to load, when to route to a faster worker, recovery paths, focused checks), while Opus handles the hard reasoning. Promotional tone, but the architecture is concrete.

@kunchenguid adds the routing map after a busy release week: interactive orchestrators (Opus 5.5, Grok 4.7), premium intelligence for escalations (GPT-6 Astra, Fable 5.1), planners (Opus 5.5), implementers (GPT-6 Sol, Grok 4.7), and trivial fixers (GPT-6 Luna), with Muse Spark 1.3 and DeepSeek V4 Flash as budget fallbacks. He feeds those preferences to Firstmate and says Jev makes the routing cheap; notably, Opus 5.5 spans nearly every bucket, so Anthropic is his pick if he can keep only one subscription.

Agent fleets are talking to each other, and grading each other

Several posts describe agents collaborating, reviewing, or messaging on humans' behalf, and the ones with evidence lean hard on verification.

@nicbstme's workplace story is the most vivid: his agent DMs colleagues on Teams for context, sometimes reaching a colleague's own agent, which digs through emails, documents, meetings, or cloud storage and sends back the answer. He frames most agent-to-agent traffic as context exchange and predicts a "company brain" holding all context with access rights attached. The human punchline: a colleague got a Teams message "from" him while he stood there holding a coffee.

@0xCodez shares a 27-minute demo attributed to SpaceXAI engineer Nick Miller, who says he runs 20+ GrokBot agents in a loop and graph, including Chief of Staff, PM, and QA bots. Secondhand and hypey, but it's another fleet-over-single-bot data point.

@muratcan's four-step loop is the boldest claim: agents write evals from failing call traces, simulated callers phone a deployed receptionist with real tools and a sandbox EHR, an LLM judge plus deterministic end-state checks score and cluster failures, and reviewer agents critique stacked PRs, one fix per root cause. He's explicitly echoing Anthropic's recursive self-improvement framing, so calibrate the enthusiasm accordingly.

The most checkable recipe comes from @mattpocockuk's summary of a talk by @poteto (who says they shipped 2,500 PRs last month): lock agents down with abstractions they can't misuse plus an internal framework called Dune, build verification infrastructure with custom CLIs and factory-ready deploys, and maintain synced "feature maps" so agents can navigate the codebase. @martin_casado amplifies @poteto running 10+ Cursor projects in parallel with pstack.

The hand-written code debate rolls on

@GergelyOrosz finds the uproar over DHH declaring the age of hand-written code over "a bit interesting," arguing it was clear since roughly January to anyone using the models and paying attention, and links a paywall-free version.

@GeoffreyHuntley pushes further: language convergence is "crossing the overton window," and languages that don't adapt to an agent consumer base will die. Why add human ergonomics like operator chaining if humans no longer write the code? His personal split: Rust where memory and performance guarantees matter, Elixir/OTP for everything else, partly because agents can attach to an actor process and troubleshoot fast.

Lighter takes round it out: @tobi built a disk-space utility and muses "You can simply wish software into existence," @thekitze jokes about mastering IT for 20 years just to stare at a screen daily, and @alexandr_wang posts satire mocking the gap between consumer AI gadgets and the demand for enterprise B2B integrations.

Infra drops: agent sandboxes, billion-token SQL, TPUs in orbit

Docker's Sandbox Kit Specification, built with the Linux Foundation, targets the exact pain the fleet posts imply: an open standard (Apache 2.0) declaring what an agent may do, where it may reach, and what it may touch.

@charles_irl introduced Quail, an open source system for AI-SQL queries, joking about crossing a query planner with an inference engine. The quoted announcement from @sh_reya says the Modal collaboration plans queries and LLM inference together to reach 1B+ input tokens per minute on a single H100 for one query. Impressive if it replicates outside their setup.

Google announced Project Suncatcher's first orbital test: a prototype satellite launching to evaluate how TPUs perform in space. And @zephyr_z9 retweeted @elonmusk listing Colossus 1 at 150k H100, 50k H200, and 30k GB200, with Colossus 2 at 110k GB200 and 440k GB300, plus another 220k GB300 teased before the post cuts off. Those are Musk's own numbers, unaudited.

Building for agents: MCP craft, bubbles, and searchable ontologies

@RhysSullivan's MCP guide is the day's most actionable post for tool builders: give the MCP full parity with your dashboard (including deletes, made safe by deep-linking into the product), skip lazy loading and codemode, ship a search-docs or skills tool, provide a deep-link-into-product tool, and stop restricting which clients can OAuth. He names PostHog and Sentry as references and argues that shipping your API spec plus CIMD OAuth covers 99% of agents.

On UX, @alexcornell explains a founding Muse decision: every message lives in a bubble because Muse "should feel like an entity not a tool," an agent rather than a search box, a position he'd argued against years earlier. @chadwahl's Palantir thread calls the Ontology an indexed space and vector store where enterprise search across links, relationships, logic, and action logs falls out of the box, and he's taking architecture questions. @deimantasnork demos a live Kanban "Captain's Deck" for Firstmate on Herdr-gpui, tracking work from Charted Next to Landed. And @sarlev_ proposes Urbit as the identity and networking layer for Omarchy to deliver an iCloud-like experience, quoting @urbit's announcement and framing it as helping @dhh kill Apple; more rallying cry than roadmap.

Practical Takeaway

If you're running agents, the strongest-supported move from today's posts is a role-based model lineup (orchestrator, planner, implementer, trivial fixer, as @kunchenguid describes) with a fast decision layer doing the routing, since @Av1dlive reports ~80% cost and time savings from exactly that. Two caveats before you commit: benchmark Jev against CLM on your own workload, because CLM's claims come from its own announcement thread, and only let fleets run where you've built verification and guardrails. Even the team reporting a closing self-improvement loop still keeps humans on merge and deploy.