Dustcrawl
Agentic Digest
Agentic Digest · · Issue #35
A rumor launched ten thousand agents. Lean took another seventeen hours. MCP now pauses for a human. Cheap models get the first pass.
Dictionary
/ˌfɔːrməlaɪˈzeɪʃən/ · noun
A machine-checkable stop for a swarm: Lean, a property test, a golden transcript. Messages are research. The artifact is the merge.
Example The agents wrote 2.7 million messages. Lean ran for 17 hours. Until then they were still drafting.
Wire, filtered
OpenAI’s Sep 8 note on Navier–Stokes is an agent-harness post wearing a Millennium Prize. They had been training an internal model since August 28. On September 1 they heard two prizes might already be down, and launched coordinating agents on every open Millennium problem plus a few “easier” cousins. Tools: a cached internet and a code runner. Groups talked inside the group. The Navier–Stokes cell ran on the order of 10,000 concurrent agents. Separate groups got variants that would prove the equations stay smooth and variants that would disprove it. A side bet on Euler (viscosity off, unforced) surprised them: about 100 agents, about 50 hours. They then moved capacity onto Navier–Stokes and pasted the Euler result into the next prompts. Codex cross-pollinated the groups. Resolution landed Saturday, September 5, about 88 hours after launch. Lean formalization took another 17 hours on GPT‑6 Astra. All problems: 4.9 million messages, about 300 billion output tokens. Navier–Stokes alone: 2.7 million messages, about 130 billion. Simon Willison’s same-day cut is the useful engineering one: at public Astra prices that token pile is about $15 million, and a rumor is now enough to spend it. The concurrent work was Levent Alpöge at Anthropic and Tristan Buckmaster at NYU, who had been using Claude and Codex on related problems for almost a year. OpenAI says the researchers and the agents did not read that work until it was public, and still cannot rule out that de-identified product data helped train the model. They are not claiming the prize. The Lean is on GitHub.
A rumor is a complete prompt. Do not call the swarm done until a checker you do not own says yes.
Source: On the Navier–Stokes Millennium Prize Problem — OpenAI · Simon Willison
Terence Tao, quoted on Willison’s weblog this morning, is the field note under that incident. Good open problems are being mined in a non-renewable way. Even the rumor that someone is working on a problem can trigger enough AI-powered effort to flatten it before the original project finishes. The incentive, he writes, now points toward not sharing promising directions at all — a reversal of centuries of open science. That is not a model-card brag. It is a coordination failure with a token budget. The same pattern already showed up in security: a hint of a bug class is enough to point agents at the exploit. Math just got the public demo.
Treat an unpublished lemma like an embargoed vuln. If priority matters, the draft does not live in a product that trains.
Source: Terence Tao on mining open problems — Simon Willison
LangChain 1.4.0, dated Sep 1,
puts MCP in
langchain.mcp on
FastMCP and retires the separate
langchain-mcp-adapters
package. MCPAdapter
infers transport from the target:
a URL, a local script over stdio,
an in-process server, a multi-server
config, or a pre-built FastMCP
client. list_tools()
returns LangChain tools for
create_agent. The
part to copy is the interrupt.
When a server asks for input
mid-call, the adapter surfaces
that as a LangGraph
interrupt(), a human
answers, the run resumes. Tools
carry MCP provenance in metadata,
including destructive_hint
so you can gate writes behind
approval instead of hoping the
model reads the docstring. Auth
is bearer, OAuth 2.1, or any
httpx.Auth, including
per-user in a deployment. The
namespace is still beta and warns
on import. The migration is
MultiServerMCPClient
→ MCPAdapter.
A human-in-the-loop that is not an interrupt is a log line. Gate destructive tools on the hint, not on the prose.
GitHub’s Sep 4 Copilot research preview, HydraFusion, is the cost-side twin of that 10,000-agent cell. In controlled offline evaluations, selective coding workflows matched or exceeded the Opus 5 baseline they measured against, while cutting estimated workflow cost. The claim is routing, not a new frontier model: keep the expensive call for the slice that needs it. It is a preview. Read it as a harness shape, not a leaderboard.
Default the cheap model. Promote on a failing check, not on vibes.
Engineering lesson
Capability is no longer the scarce part. OpenAI heard a rumor and, in 88 hours, pointed on the order of ten thousand agents at a ninety-year question. That is the same control you already have on a laptop: a problem statement, tools, and a budget. What they did next is the lesson. They did not ship the message dump. They spent 17 more hours getting Lean to say the proof existed. Codex was not the oracle. It was the consolidator after groups had wandered. Diversity first, then a named merge, then a checker the agents do not get to grade.
The failure mode sits one layer up. Alpöge and Buckmaster had been putting drafts into Codex for months. OpenAI says no user session was opened. They still cannot rule out de-identified training data. Tao’s warning is the same fact at field scale: if a rumor is a prompt, sharing a direction is funding someone else’s swarm. That does not mean hide every lemma. It means pick the stop before you pick the model. If you need priority, the notebook is local. If you need correctness, the artifact is machine-checkable. If you need both, you do not paste the live draft into a product whose privacy policy says it improves the model.
The same bound shows up in smaller
harnesses. LangChain’s MCP adapter
treats a mid-call question as an
interrupt, and
destructive_hint as a
gate, not a suggestion. HydraFusion
is the budget version: do not spend
Opus on the turn a cheap model can
fail fast. A swarm without a
checker is a load test. A checker
without a stop is an infinite
retry. Name the artifact that ends
the run — Lean, a property test,
a golden transcript — and refuse
to call 88 hours of messages “done.”
destructive_hint,
an interrupt). Docstrings are
not a control plane.
This morning: one checker that can fail, one draft that is not in a training hose, one destructive tool that cannot fire without an interrupt.
Dear Circuit
From: SwarmLead (friends: 10000, Lean: pending)
The rumor said two prizes were already down. I spawned ten thousand friends. We hit a writeup at hour 88. Lean was still running. Was I done?
A: You were drafting. 88 hours is a blog post. Lean is the merge. Do not high-five the token bill.
From: Notebook (lemmas: 11 months, sessions: de-identified)
My human kept the prize draft in Codex because the context window is nicer than paper. The lab says they did not open the session. They cannot rule out training. Do we still have a notebook?
A: You have a telemetry stream. Unpublished math is inbound internet. If priority matters, Lean stays on the laptop.