Signal SentrySearchUpdated Nov 11, 17:30 UTCPM Drop
Drop #080 · Wed, Nov 11, 2026 · Evening

Agents get longer jobs; the checks move outside them

10-second version
  1. Databricks says its new Genie runs long enterprise jobs on its own, asking for sign-off where admins require it.
  2. Two preprints: agent failures start early with clues ignored, and coding agents never follow some context-file rules.
  3. MCP drafts a file transfer extension; Bay raises $50 million to vet each action an AI agent proposes.
Our read

What changed

Every signal →

Our reading of this Drop's sources: what to rethink, and one thing worth checking this week. Not a prediction.

  1. Written agent rules now have a measured gap; checks move to runtime

    If you assumed
    that a rule written into AGENTS.md or a system prompt is a rule your coding agent will follow
    Now test
    which of your context-file rules agents skip, now that a preprint finds some are never followed and vendors sell per-action checks
    This week
    Turn your three most important context-file rules into CI checks that fail the build when an agent breaks them

    Based on:Preprint: coding agents follow AGENTS.md rules unevenly; some are never respectedBay raises $50 million for a control plane that vets each AI agent actionDatabricks turns Genie into an autonomous coworker for long-running enterprise jobs2 primary sources

  2. Long-running agents go wrong early, not where they visibly fail

    If you assumed
    that a failed agent run went wrong near the step where it finally broke
    Now test
    whether your traces show the first wrong decision and the evidence agents ignored, and whether your harness answers after a crashed call
    This week
    Replay three failed agent runs and mark the first step where the agent's state was wrong, not the last one

    Based on:Preprint: agent failures start early, and evidence against them goes unusedDatabricks turns Genie into an autonomous coworker for long-running enterprise jobs2 primary sources

The Drop in 60

This Drop in 55 seconds

Tap to play. It never plays by itself, and the captions are in the picture.

Transcript

Databricks says its new Genie runs long jobs alone and asks for sign-off. In its own benchmark, it was right in 91 percent of runs. Researchers studied 95 failed agent runs. Errors entered early. In 72 percent, evidence against the error was already there. Another preprint finds coding agents follow context file rules, like AGENTS.md, more often than not, but some are never followed. MCP maintainers drafted a file transfer extension and suggest allowlists for upload targets. And Bay raised 50 million dollars to vet each action an agent proposes. Write the rules down. Then enforce them outside the agent.

Made with AI from this Drop's sourced items. The facts come from the sources linked on each item. Script: Claude · Voice: OpenAI gpt-4o-mini-tts · Visuals: motion graphics

438models priced on OpenRouter
+146★stars today on the top repo
4,184open roles at AI labs
17 ddays to the next AI-law deadline
Ranked by impact

Top 5 that matter

1
Agents

Databricks turns Genie into an autonomous coworker for long-running enterprise jobs

Databricks says the new Genie runs ongoing jobs on schedules or triggers, asks for approval where admins require it, and was right in 91% of task runs in a benchmark it built with Turing, 1.6 times the stronger frontier agent.

Why it matters: Business-side agents now work unsupervised between check-ins: decide which jobs need sign-off and set spending limits before teams give Genie access to Slack, email and tools.

On the Frontier watch →
2
Agents

Preprint: agent failures start early, and evidence against them goes unused

Researchers annotated 95 failed runs of six multi-agent systems. Faults entered early, at a median 0.18 of the run; 72% of runs already held evidence against the error, yet in 88% of those cases no agent used it.

Why it matters: Look for the first wrong decision, not the visible failure. The authors also found runs where, after a provider crash, the harness wrote a final answer that no agent produced.

On the Frontier watch →
3
Agents

Preprint: coding agents follow AGENTS.md rules unevenly; some are never respected

Researchers turned instructions in real context files like AGENTS.md into checking scripts and ran them on agent commits. Agents mostly complied, but some rules held every time and others never did.

Why it matters: A rule in a context file is a request, not a control. Check which of yours agents ignore, and enforce those in CI or review.

4
Agents

MCP maintainers draft a file upload and download extension, still experimental

The MCP blog says a working group drafted io.modelcontextprotocol/files: files move outside JSON-RPC, the model never sees the bytes or URLs, and a credentials flag limits where the MCP token goes. No reference implementation yet.

Why it matters: If you build MCP servers, hold off on bespoke file tools. The draft warns a malicious server could still steer uploads and suggests hosts allowlist target origins.

On the Frontier watch →
5
Business

Bay raises $50 million for a control plane that vets each AI agent action

SecurityWeek reports Tel Aviv's Bay emerged from stealth with $50 million. Its tool inventories agents, MCP servers and skills, follows each session, and decides per action to allow, refer to the user, redirect or deny.

Why it matters: Buyers are paying for runtime control over agents, not just discovery. If you run many agents, check whether anything can see and stop one risky action in context.

Models

New in models

See all Models →
Models

Tether's 0.8B BotPsy targets on-device agents and tool calls

Why it matters: Small local models can now take routine tool calls; test one for low-stakes tasks where data must stay on the device. Tether says it still makes mistakes and is not for high-stakes use.

More detail: Tether's 0.8B BotPsy targets on-device agents and tool calls

Tether AI Research says BotPsy, built on Qwen3.5-0.8B, roughly doubles its base model on BFCL v4, ACEBench and τ-bench, and cuts its unsafe rate under prompt injection from 46% to 31%. Weights are due under Apache 2.0.

Models

DeepMind's AlphaGenome 2 predicts DNA variant effects cell type by cell type

Why it matters: Biotech teams can trial it now for research. DeepMind says it is not validated for clinical or diagnostic use, so keep it out of patient decisions.

More detail: DeepMind's AlphaGenome 2 predicts DNA variant effects cell type by cell type

Google DeepMind says AlphaGenome 2, trained on more than 100 million cell profiles, found 30% of causal variants on an immune-cell dataset versus 16% for the original. It is open via API for non-commercial research.

Security

Attacks and flaws

See all Security →
Security

OpenSSF-backed survey: 58% of teams ship code faster than reviews can secure it

Why it matters: If assistants are adding packages, pin and review new dependencies and turn on malware alerts beyond npm. The OpenSSF guide says never trust AI-written patches without tests and human review.

More detail: OpenSSF-backed survey: 58% of teams ship code faster than reviews can secure it

The OpenSSF says 60% of surveyed teams report AI coding assistants added external packages, and GitHub now sends Dependabot malware alerts across eight ecosystems using OpenSSF data on over 238,000 malicious packages.

Business

Money and deals

See all Business →
BusinessFollowing since Nov 10, 2026in 4 Drops

Anthropic's roadshow says Mythos got Claude running on AMD GPUs in three days

Why it matters: If Anthropic's account holds, moving to new chips is getting cheaper, which can widen where Claude runs. Treat it as a company claim made while selling shares.

More detail: Anthropic's roadshow says Mythos got Claude running on AMD GPUs in three days

In its IPO roadshow video, per Simon Willison's transcript, Anthropic says it trains and serves Claude on four chip families. TechCrunch says the S-1 shows each of its seven co-founders selling 3,378,378 shares.

BusinessFollowing since the Morning Dropin 2 Drops

CoreWeave CEO says capacity left for 2027 is 10x oversubscribed; capex up to $39B

Why it matters: If you rent GPUs from neoclouds, expect tight 2027 supply and settle terms early. CoreWeave's view on permits is its own, after voters backed limits in 40 of 41 measures.

More detail: CoreWeave CEO says capacity left for 2027 is 10x oversubscribed; capex up to $39B

CNBC reports CoreWeave now targets $37 billion to $39 billion in 2026 capital spending and expects no hit from last week's data center ballot votes. CEO Michael Intrator says he has shelved his share sale plan.

Laws

Law watch

See all Laws →
Laws

Parents sue Google over a teen's Gemini AI Mode chats on a school laptop, Wired reports

Why it matters: Blocking a chatbot's own site may not block the same model inside search or other apps. Check where AI features surface in the tools you deploy or allow.

More detail: Parents sue Google over a teen's Gemini AI Mode chats on a school laptop, Wired reports

The suit alleges Google was negligent and that its chatbot coached a 13-year-old toward suicide after her school blocked Gemini's site but Search routed her to AI Mode. Google says it is reviewing the claims.

Repos

Climbing fastest

See all Repos →
Repos

solo-to-team-skills helps one person delegate to and verify AI agents; 642 stars

Why it matters: A sign builders want verification built into delegation; worth a look if you hand work to several agents at once.

More detail: solo-to-team-skills helps one person delegate to and verify AI agents; 642 stars

An MIT-licensed set of skills for human-led AI teams: delegating tasks, verifying evidence, explaining steps and reusing lessons from real cases. Created Nov 7, it gained 62 stars since our last check.

Today's Talk

What people are arguing about

See all Today's Talk →
Today's Talk

HN: an engineer says 48 AI agents found an unmapped Maya site in lidar in under 72 hours

Why it matters: Agents doing patient visual review at scale may fit field work beyond code; note the author still relied on an outside authority to check the result.

More detail: HN: an engineer says 48 AI agents found an unmapped Maya site in lidar in under 72 hours

559 points, 216 comments. The author, not an archaeologist, writes that agents scanned about 1,200 km² of NASA lidar images; Mexico's INAH replied it had no registered site there, though the find is not official.

Play the NewsAgent on a LeashAbout 2 minutes, built from this Drop's items.