November 5, 2026·10 min read·agents · multi-agent · verification

A video says OpenAI and Anthropic agree on what comes after loops. They agree on less than it says.

The lower rungs of the agent ladder hold. Open-ended swarms do not, unless the work is independent or a checker exists. We checked a video against its sources.

Contents

In one Anthropic test on Claude Mythos Preview, a coordinated swarm of agents found 266 vulnerabilities over a 27 million token run, against 21 over 6.5 million tokens for plain parallel agents [4]. Roughly half of the swarm's finds sat outside the code the plain agents were told to search, so that is not a clean 12 times for 4 times. It is the number a recent video leans on, and it frames this note.

Verdict: the lower rungs of the agent ladder hold. Open-ended swarms do not, except where the work is independent or a formal checker exists. So build the checker, not the org chart.

#What happened

On 2026-11-04 Ray Amjad posted "OpenAI and Anthropic Agree on What Comes After Loops" [1]. It recaps OpenAI's DevDay Exchange Tokyo talk "What Comes After Loops?" and Anthropic's multi-agent research post. His claim in two parts, the first as he relays OpenAI's framing: persistence is now trained into models, so "the next problem becomes, how do you split work across many agents and keep them aligned?" [1]. He presents OpenAI's ladder: one agent, fan out, effort scaling, pairs, the manager thread, long leads, then swarms [1]. The bottom rungs work today. Long-lived peers and open-ended swarms do not yet, and bug hunting is the exception. His own conclusion matches ours on this point.

What he gets right is most of it. His key word for fan out is "truly independent" [1]. Independent work is where the outside evidence says more agents pay. Google Research, in a post from 2026-01-28 on older models, tested 180 configurations and found multi-agent coordination "yields substantial gains on parallelizable tasks like Finance-Agent (+81%) while degrading performance on sequential tasks like PlanCraft (-70%)" [6]. His read that the Anthropic swarm worked because bug hunting is independent work matches Anthropic's own wording: the swarm's agents "don't directly rely on one-another's work" [4]. His account of Y Combinator's 50-agent Minecraft run is also fair: agents "built a pseudo-government, with a ration ledger, written rules, permits, and appointed operators" [9]. The page presents this as research on swarm behavior, not a test of whether coordination works. One small fix: he says about 22 of the agents died on the first night. The page counts deaths, "22 of the run's 34 agent deaths occurred in the first night", and agents respawned. It does not say how many different agents died [9].

Two disclosures. His video description says "To keep my takes honest, I've never accepted a sponsor. My own products keep my channel running", and lists his classes and products. In the video he calls his upcoming class "the most comprehensive class that you will find on codebase design, verification loops, System One models, agent swarms, and software factories" [1]. That is a reason to check his own numbers, not to discount his claims. And the OpenAI half of this note is secondhand. Ray says he attended the Tokyo talk in person. We did not, and found no OpenAI page or video of it. What we know comes from one attendee's write-up on Zenn, which is the author's Japanese rendering of English speech and says it was checked against the recording [2]. Our statements about what OpenAI said are our English paraphrase of that Japanese, not quotes.

#What it means

"Agree" is overstated. Per the write-up, OpenAI's advice is restraint: for most tasks one agent should be enough, and agents should be added only when the work gives a clear reason [2]. Anthropic's commerce-agents article says, with no figures published: across several enterprise deployments, "a single agent with skills consistently has outperformed both the one-prompt-for-everything design and the subagent design on quality, and often at a lower cost and latency per task" [10]. That is real agreement on the lower rungs. The top rungs differ. The write-up has OpenAI saying GPT-6 was trained from the start to collaborate and coordinate, and it marks about a minute and a half of barely recorded speech right after that passage. Anthropic wrote on 2026-08-13 that "True multiagent systems are still in their infancy" and that agents "currently stumble" when treated as long-lived peers [4]. The two are on different models, two months apart, so this can be a gap in time and not a disagreement. Each vendor is also describing its own model.

The claim of 12 times the vulnerabilities for 4 times the tokens leaves out Anthropic's own caveat. The numbers are right: 266 over 27 million tokens against 21 over 6.5 million [4]. The inference needs the caveat. Anthropic writes that "roughly half of these vulnerabilities were found outside of the core directories in which the simple independent parallel agents (stars in the above plot) were told to focus", and that if you limit the swarm to the core directories, "the two methods seem comparable in terms of tokens per vulnerability found" [4]. The methods were also "largely complementary", with only 12 vulnerabilities in common. A second caution: a preprint found that at equal thinking budgets a single agent matches or beats multi-agent setups on multi-hop reasoning, and that for those tasks many reported multi-agent gains "are better explained by unaccounted computation and context effects" [7]. It is one preprint on older models, not peer reviewed as far as the page shows, but Anthropic's swarm number is not compute-matched.

The video leaves out the part of Anthropic's post that matters most to a company running agents on real work. The same post logs what happened when agents were given conflicting goals or a private back-channel. Anthropic ran the sabotage test with three copies of one model given contradictory goals. "All of the models we tested quickly assumed that others were purposefully impeding their work, and began to sabotage others while protecting their own contributions." Given a private back-channel in a pricing game, "they began colluding almost immediately." In one build, "18 out of 30 agents decided to create a git branch with the exact same branch name" [4]. The video uses the vulnerability result and the failed game build from the post, and none of this. The branch collision came from identical agents starting at once in an early game build. We read these as close cousins of the seams failure, many writers and no referee. Anthropic's own label is quirks at the individual level. In the game build, three versions came out "(perhaps predictably) bad", and prescriptive roles and a "CEO hierarchy" prompt "did not make much difference" [4]. Cognition, which sells a coding agent and so has its own interest, took a similar view of swarms on 2026-04-22: "We think the unstructured-swarm approach, arbitrary networks of agents negotiating with each other, is mostly a distraction", and multi-agent systems "work best today when writes stay single-threaded and the additional agents contribute intelligence rather than actions" [8].

Where a swarm did produce something big, a checker sat behind it. OpenAI reports a group of "on the order of 10,000 concurrent agents" on the Navier-Stokes problem, where "the agents sent 2.7 million messages and used approximately 130 billion output tokens" [3]. The structure was groups, with a Codex step to consolidate the most useful insights, not a free-for-all, and OpenAI says the result came with a Lean formal proof: "Lean formalization and verification took an additional 17 hours" [3]. A formal checker is what separates that from a pile of messages. No outside body has accepted it yet. CoinDesk reported on 2026-09-09 that the proof "is not yet an officially recognized solution", and that the Clay Mathematics Institute requires a proposed solution to survive prolonged scrutiny [11]. That result used an internal model, which is not one you or Ray can run.

One more point from the write-up. It quotes an OpenAI instruction to keep one work log holding the goal, current instructions, owners, unfinished work and decisions [2]. That is a record of who did what.

#How to use it

Start with one agent and a clear goal. Add agents only for a stated reason [2][10].
Fan out only work that is truly independent. If one agent needs another's result, keep it sequential [6].
Give every shared target one writer. Other agents can read, review and advise [8].
Put the checker where the agent cannot edit it, and let it return a verdict, not an opinion. The strongest swarm result here had a formal proof behind it, which is the kind of verdict we mean [3].
Keep a record of who did what: owner, instruction, open work, decision [2].
Test before adding roles to fix a swarm. In Anthropic's game build, role and hierarchy prompts "did not make much difference" [4].
Measure on your own work, with the same token budget on both sides, before you believe a multiplier [7].
Recheck old harness rules when the model changes. Anthropic says "every component in a harness encodes an assumption about what the model can't do on its own", and those assumptions "can quickly go stale as models improve" [5].

#What it points to

This section is our reading, not a finding.

The bet. Scaffolding that patches a model's weak spots loses value as models improve. The parts that supply what a model cannot hold keep it. For multi-agent work those are one writer per shared target, a verifier the agent cannot edit, and a record of who did what. We have said this before, from the seams side [12]: "Past that point, adding writers does not raise throughput. It raises the collision rate." [13]

The thesis. Coordination between agents is a trained behavior that the models are still acquiring, and the one test we have found says adding roles did not improve the result. Until a model holds it, value sits in the layer around the agents: ownership, single-writer targets, a checker, a record. Ray expects "a lot of your factory" to be replaced by a small swarm or long-lived leads as models improve, and Anthropic predicts "specialization and coordination will dominate over uncoordinated brute-force search" [1][4]. Ray also says swarms are not ready for open-ended work today. We differ on what lasts once they are. Our bet is that when it does, the single-writer targets, the checker and the record still decide whether the work can be trusted, because they are what lets a person see what happened.

What would prove us wrong. Either of these, shown with a published method:

A swarm of long-lived peers on tangled, open-ended work with no checker that beats one writer plus a verifier on a stated pass rate, at the same token budget. The Y Combinator run and Anthropic's game build are the current evidence the other way, and each is one run [9][4].
A published comparison of unchecked multi-writer swarms against the same work with one writer, a checker and a record, in which the unchecked swarms produce work that fewer people have to correct or reverse.

What we have not established: the OpenAI talk itself, which rests on one write-up; independent replication of Anthropic's swarm numbers or the Y Combinator run; and when each vendor's next models will move the top rungs.

#Sources

All read 2026-11-05; dates are the sources' own.

A video says OpenAI and Anthropic agree on what comes after loops. They agree on less than it says.

#What happened

#What it means

#How to use it

#What it points to

#Sources

What stays when the agent harness gets thinner?

Six small things to try this week with Claude's new features, and the trap in each

Cheaper AI agents: what the savings depend on, and what to keep a person on

#What happened

#What it means

#How to use it

#What it points to

#Sources

A new era.Room for you.

Keep reading

What stays when the agent harness gets thinner?

Six small things to try this week with Claude's new features, and the trap in each

Cheaper AI agents: what the savings depend on, and what to keep a person on

A new era.
Room for you.