DeepMind’s AI Co-Mathematician landed earlier this month. Twenty-three pages, no ablations, and a fresh state-of-the-art on a benchmark whose name you have already, quietly, begun to forget. On a first read I was ready to file it under “another agentic-math demo”, a genre now large enough to have its own NeurIPS workshop. But there’s a single sentence about three pages in that’s worth the price of admission:
Rather than silently restarting or scrubbing the history of these unsuccessful attempts, the architecture treats dead ends as first-class, permanent outcomes.
That, not the orchestration diagram, is the paper.
The premise the paper assumes
The paper never argues for why an AI mathematics workbench should exist, it assumes you’ve already been sold. Pushmeet Kohli, its senior author and DeepMind’s VP of Science, made that case separately, in a recent Daedalus essay. One line from it explains a design choice the paper itself presents almost in passing:
as these hypotheses-generating AI agents become more powerful, it is increasingly important for human scientists to be extremely precise in specifying the problem they want these agents to solve. Otherwise, the AI agents tend to find a solution that violates assumptions the human scientists were making but had not included in the original problem description.
That’s why the Project Coordinator never starts solving directly. It first opens a dialogue to refine intent. Not UX polish, but the engineering response to a documented failure mode. Read this way, the paper is a position statement: one instance of how, with the why argued elsewhere.
The architecture
The architecture is the standard 2025 multi-agent shape. A top-level Project Coordinator dispatches workstream coordinators for parallel lines of inquiry (literature search, library development, theorem proving), each of which delegates to specialist sub-agents: a search agent, a coding agent, Gemini Deep Think as a proof verifier. Reviewer agents gate every sub-agent’s output before reports flow back up the chain. Organizationally it is a medium-sized consulting firm that exists for forty minutes and then evaporates. You can sketch it on a napkin and it’ll look like every agent paper this year.
The paper is explicit that this is a harness, not a model:
the AI co-mathematician is a harness designed to be used with standard, commercially available language models, and does not rely on any custom model behavior or training.
The headline number is 48% on FrontierMath Tier 4. Base Gemini 3.1 Pro alone scores 19%. The 29-point delta comes entirely from the scaffolding. That’s the central claim of the paper, and arguably of 2026 in general: progress is increasingly coming from the harness, not the model.
Failure as durable project state
However, keeping a record of what didn’t work is not a new idea in 2026. Plenty of coding agents in production already maintain a progress.md or a NOTES.md, logging what they tried and why it broke. The Co-Mathematician’s own paper is candid about the lineage. It cites Claude Code, Codex, and Antigravity by name, and borrows the practices that make them work: design docs to keep agents on a path, test pipelines for verification, version control to track state. So the question isn’t whether this system logs failure. It’s why its version of the idea does more than a scratchpad file does.
Three things separate them.
First, the failure record is permanent project state the system is built around, not a note the next agent may or may not consult. When a workstream ends in failure, the architecture refuses to scrub it:
Rather than silently restarting or scrubbing the history of these unsuccessful attempts, the architecture treats dead ends as first-class, permanent outcomes.
Preserving this ‘negative space’ of the project’s history is crucial for tackling difficult mathematical problems.
The Project Coordinator reads that negative space before dispatching the next workstream. A progress.md is a file the next agent might open if it thinks to; here, consulting the failure record is wired into the loop. The coordinator is designed to build its next attempt out of the recorded failures rather than rediscover them.
Second, and I think this is the underrated part, the domain is mathematics, where the final artifact is LaTeX, and that turns out to matter more than typography. A failed lemma is a specific object: this bound, this approach, ruled out for this reason. Math gives you failures with sharp edges, and sharp-edged failures are the ones worth keeping, because the next agent can actually reason against them instead of just acknowledging they happened. The structured artifact pulls the failure record toward being structured too. (The paper is honest that this cuts both ways: mathematicians read a well-typeset document as a rigorous one, while the models “excel at flawless LaTeX” and “struggle with rigorous verification.” A clean proof of a false statement is still false.)
Third, a remembered dead end is cheap. Once catalogued it’s a short workspace note, not a 50,000-token transcript to re-summarize, so the coordinator’s context cost falls as the project’s history grows rather than rising with it.
All of which is simply how a mathematician already works. The notebook has the crossings-out in it. Tuesday’s failure shapes Wednesday’s idea. One early user said the system’s main value was letting him reach a dead end faster: “I could have easily spent a week dreaming about what was there, but instead I just moved on.” The intuition for “this won’t work, I’ve tried something like it” is built out of remembered failure. An agent that wipes the slate at every context boundary isn’t helping with this; it’s the colleague who arrives each morning and confidently re-proposes the thing the room collectively ruled out the previous afternoon.
It’s worth saying what the closest prior system didn’t do. DeepMind’s own AI co-scientist, the multi-agent research partner built around a generate-debate-evolve “tournament” that ranks hypotheses against each other, is designed to surface and refine the winners. The losing hypotheses are selected against, not remembered. The Co-Mathematician’s bet is that the losers are part of the project, not something to discard once they’ve lost.
The output is a LaTeX artifact
Each workstream produces a compiled LaTeX writeup as its main artifact. Not a transcript, not a chat log, but a document built like an actual research paper: exposition, results, and references. Every margin annotation links a claim back to the workspace it came from.
The output of a math research session is supposed to be a mathematical artifact, not a transcript that the user has to re-derive a paper from. Margin notes that trace each claim back to its source are the cheapest version of “make the output reviewable,” and they fall out almost for free once the workspace is already structured.
A statement, not a study
There are no ablations, and the benchmark runs use more compute than the baselines, so the 19% → 48% jump on FrontierMath Tier 4 can’t be pinned to any single component: reviewer agents, failure-tracking, workstream parallelism, the LaTeX discipline, or just “running Gemini 3.1 Pro for a lot longer.” The choice to publish the overall design rather than a component scorecard comes from one of the most rigorous labs in the world, and I read it as part of the argument: the claim is that the design is what matters. The paper is also unusually honest about its failure modes, like reviewer agents that converge on flawed-but-undetectable proofs, or review loops that degrade into “death spirals.” That kind of disclosure is what labs with something to hide tend to leave out.
There is one concrete real-world win: Marc Lackenby used the system to resolve Problem 21.10 of the Kourovka Notebook, an open problem in group theory. The first write-up was flagged wrong by a reviewer agent, but it contained, in Lackenby’s words, “a really, really clever proof strategy,” and the reviewer’s own critique was specific enough that he saw how to close the gap. He did; the system wrote up the corrected proof. Credible, and a nice illustration of failure-as-input, though human-in-the-loop the whole way through.
The harder problem underneath: episodic memory
The Co-Mathematician’s record-keeping is a clean demonstration, but it points at a much harder problem sitting just underneath. Capturing failure is the easy half. Carrying it is the hard half.
Run a long enough research project and you don’t have one failed attempt, you have a hundred. How do a hundred dead ends fit in a context window? They don’t. A full transcript of each is out of the question, and a one-line summary throws away the specific detail that made the failure informative in the first place. And the detail is the whole point: some of the best results in science arrive only after hundreds of failed attempts, where the failures aren’t noise around the signal so much as the slow triangulation toward it. An agent that can hold three failures in working memory and a human who remembers a decade of them are not doing the same job.
And remembering is only the floor. The harder thing a mathematician’s memory does is hold a dozen old dead ends side by side and notice that two of them, each a failure on its own, combine into the approach that finally works. That cross-pollination, the slow collision of half-finished ideas that eventually throws off a spark, is most of what we mean by insight, and it depends on the attempts being available all at once rather than retrieved one at a time. Human memory makes this feel free: ask yourself what you did last Tuesday, or which approach you gave up on last year, and the answer is simply there, ready to be recombined. For these agents almost none of it is. If it isn’t in the context window, it effectively never happened.
So the real open problem isn’t “should agents log failure.” That’s settled; everyone does. It’s memory, and specifically episodic memory: a durable, queryable store of specific past attempts (what was tried, in what context, and why it failed) that survives across sessions instead of living and dying inside a single context window. The hard part is retrieval, compression, and recombination: surfacing the three relevant failures out of a hundred without either blowing the context budget or flattening each one into a one-line summary that has already thrown away the detail that made it worth keeping, and then holding enough of them in view at once that they can actually be combined. The Co-Mathematician is taking the first step in this direction: it keeps its history inside one workspace, one project, broadly one sitting. The version that matters carries it across all of them, and eventually does what a researcher’s memory does: it consolidates years of specific dead ends into something more like intuition, the quiet shift from recalling each failure to simply knowing better. Episodic memory of what didn’t work, done well, is, I’d wager, going to be one of the defining shifts in agentic systems over the next few years, more so than any particular orchestration diagram, this one included.
The takeaway
If I had to extract one thing this paper changes about how I think about agent orchestration, it would not be “DeepMind has a new SOTA on FrontierMath.” It’s two things, really, pointing the same way. The first is that the failure log is the new context window: long-horizon agents should treat dead ends as part of the project’s persistent state, not as noise to clear out at session boundaries. The second is that this is the thin end of a much larger wedge: a durable failure record is the first primitive of a real agent memory system, and that record, not the next clever model, is where the next bit of headroom lives. Plenty of systems now jot failures down; almost none are built to use that record the next time around, to surface the relevant attempts and recombine them, let alone across sessions. The Co-Mathematician’s bet is that they should.
Kohli’s framing in Daedalus was that humans are becoming architects of questions rather than solvers of puzzles. If that’s the direction, the agent system below the architect needs to remember what’s already been tried: not just the plan for what to build, but the record of everything that didn’t hold. Whether or not the rest of the architecture turns out to be the right design, that single choice is going to show up everywhere.