What lab updates matter for an autonomous AI workforce

Durable sessions, delegated work, isolated execution and approvals are the lab updates that matter. Together they do not yet make a reliable organization.

The numbers

  • 56.4%

    GPT-6 Sol on Agents' Last Exam at maximum reasoning effort

    A vendor number

  • 128

    agents the ScaleMAS preprint says it scales up to, on four agentic tasks

    A preprint. It does not show that 128 agents can run a business

  • 75 to 94%

    of eligible traffic from the strongest models judged monitorable

    Estimate from a METR and Redwood Research pilot red team of OpenAI’s own monitors. Not a reliability figure

Show all figures
  • 60%

    lower cost per task OpenAI claims against Claude Opus 5, with a higher score

    A vendor claim

I asked for a sweep of official documentation, lab papers and repositories, and for everything from that day that matters if you want AI workers to do the work of an organization. I wanted one pass over many labs instead of following each release note myself.

The result covers OpenAI, Anthropic, Google, Microsoft, LangChain and AWS in some detail. Academic research and work involving physical machines get less attention.

What I found

  • The useful changes are in infrastructure. Official releases now ship durable sessions, delegation that runs in the background, reusable skills, isolated execution, external authorization and production evaluation. Nothing in the study shows that putting them together runs an organization without accountable humans.
  • Anthropic’s Managed Agents split the durable session, the agent loop and the sandbox into separate parts, and are built so that the tokens are never reachable from the sandbox where generated code runs. A worker can be thrown away while the record of what happened stays.
  • Approval is getting more exact. Microsoft’s Agent Framework releases tighten approval handling, for example by rejecting duplicate approval request IDs (October 2026). LangChain added an opt-in mode in which each tool call that needs approval asks for its own, so one approval covers one exact action (version 1.5.18, October 15; batched approval stays the default).
  • The model behind a name can change. Google routes requests from a deprecated Gemini model to a newer one automatically (Gemini 3.7 Flash to 3.8 Flash since 8 October), and OpenAI’s system card separates the October GPT-6 Sol and Luna from the September versions still used in Codex and ChatGPT Work. The study advises tracking the exact model, API version and surface.
  • Research on coordination points away from long agent chats. The ScaleMAS preprint has agents share work through a common store of linked messages, and says it keeps scaling up to 128 agents across four agentic tasks (submitted to arXiv on 12 October 2026). These are benchmark results, and they do not show that 128 agents can run a business.
  • The benchmark numbers come from the vendors. OpenAI reports GPT-6 Sol at 56.4% on Agents’ Last Exam at maximum reasoning effort, a set of long-horizon, economically valuable tasks, and says it beats Claude Opus 5’s best score there at 60% lower cost per task. An organization also needs its cash, promises and schedules to agree, and a finished task does not show that they do.
  • METR and Redwood Research published a pilot red team of OpenAI’s internal agent monitors (exercise from 21 July to 11 August 2026). They estimated that 75 to 94% of eligible traffic from the strongest models was monitorable and found ways an agent could evade the monitor. OpenAI says most issues are patched; METR has not checked that. The range does not say how reliable the monitors are. The work is a reason to test whether agents can evade monitoring.

The study proposes making models and workers replaceable, and keeping obligations, authority, business state and verification durable and explicit. The proposal has not been tested.

How it was made

I asked Perplexity Deep Research on 24 October 2026 and it showed 416 sources. The AI gathered the material and drafted the study. I asked the question. Vendor release notes change quickly, so treat every version and price as true at that date only.

Sources

The question
Which recent releases, papers and official documents from AI labs and open projects help build an AI workforce that can run work inside an organization?
How it was made
Perplexity Deep Research, 416 sources.