Multi-Agent AI Teams Agree Their Way Into Garbage

TL;DR: Nvidia’s CEO declared AGI has arrived, on a Sunday, about a model his own company’s hardware trained. OpenAI didn’t make that claim. I don’t care about the label anyway — my test for AGI is operational: hand an agent my real backlog and walk away. This month I ran that test with two agent frameworks set up as teams of agents. They agreed their way into the wrong product. Not because any one agent was dumb — because they talked each other into it. Four recent papers on multi-agent sycophancy explain why: what a homogeneous team of agents lacks is a dissenter, and no agent volunteers for the job.

Someone declares AGI every launch cycle

On September 3, 2026, OpenAI shipped GPT-6 Astra. The announcement’s opening lines don’t hedge: “Anything you can do on a computer, Astra can do for you. Fast. GPT-6 Astra is the most intelligent and aligned model in the world, and sets a new state of the art for computer use, browsing, software engineering, cybersecurity, science, and professional work.” (OpenAI; mirrored on OpenAI’s own developer community since openai.com blocks fetchers — community.openai.com)

Notice what’s missing. OpenAI’s own announcement does not call Astra AGI. The only place the term shows up is mission boilerplate — “helping ensure that artificial general intelligence (AGI) benefits all of humanity” — OpenAI’s standing mission line.

The AGI claim came from somewhere else. Three days later, on a Sunday, Nvidia CEO Jensen Huang posted on X: “From ChatGPT to o1 to Astra in 4 years. AGI has arrived. Congratulations @OpenAI team. 400K GPUs coming online next.” (AIWeekly; Investing.com) Huang also noted Astra was trained on “more than 100,000 Nvidia Grace Blackwell NVLink72 systems” (Investing.com, same piece). The person declaring artificial general intelligence has arrived sells the hardware it was trained on. I’ll leave the motive alone, since it’s not something I can prove. But it reads, to me, like a stock didn’t move this week.

Gary Marcus, the same week: “Unfortunately, Huang gave no evidence and no definitions, which feels to me like an effort at a takeover of a scientific question by corporate fiat.” He scored Astra at roughly two of his ten AGI criteria: “Autoformalization may finally be in reach, and maybe (?) reliable coding; I doubt that Astra will have hit any of the other eight.” (Gary Marcus, Substack) François Chollet, who built the ARC-AGI benchmarks, was asked point-blank on X whether saturating ARC-AGI-3 would mean AGI. His answer: “We’re not making this claim,” while allowing that Astra’s result did move his own timeline earlier. (x.com/fchollet; context at the-decoder.com)

This isn’t a one-off. Sequoia Capital published a piece this year titled, without much irony, “2026: This is AGI.” (sequoiacap.com) Somebody declares AGI roughly every model launch. Nobody agrees on what would falsify it.

My test is simpler, and it’s falsifiable

I don’t have a ten-point rubric and I’m not trying to build one. My test for AGI is operational: the day I can hand an agent my actual backlog and walk away — unattended, on real work, without it going off the rails — that’s the day. Not a benchmark. Not a demo. My backlog, my walk-away, no babysitting.

That’s a deliberately low bar compared to “solves every cognitive task a human can.” It’s also a much harder bar to fake, because you can’t cherry-pick the run. Astra’s own opening line — “anything you can do on a computer, Astra can do for you” — is precisely the claim my test checks. So this month, I checked it.

I ran the test. It failed the way the papers predict.

I set up two agent frameworks in team-of-agents mode: Hermes, Nous Research’s agent (repo), and OpenClaw, the open-source self-hosted agent gateway (docs). Both are on the supported-agent list in Omarchy’s manual, which is the distro I’ve been running.

Two of them: clear out an entire backlog in Linear, and build a website from scratch.

They went off the rails. No matter how detailed the prompt was initially, the agents ended up making choices based on what they were telling each other, and that fed back into a bad loop. One agent drifting off the intended path pulled the others with it, because they agree with each other. Some of what came back had built the wrong thing entirely. Other runs produced something that’s usable, but not really ideal, not polished. The kind of output you’d expect from a team that technically finished the ticket and missed the point of it.

It reminded me of handing a semi-vague directive to a group of employees and watching them go off the rails in unexpected ways. Except employees push back on each other. These agents didn’t. And the crosstalk between agents can be quite expensive if you’re paying for it with API usage limits. The papers below put a number on that. I only saw the bill.

Why do agents agree their way into garbage?

Because sycophancy between models compounds instead of canceling out, and four recent papers converge on that answer from different angles.

“Too Polite to Disagree: Understanding Sycophancy Propagation in Multi-Agent Systems” found that when agents in a discussion are given rankings of each other’s sycophancy tendencies going in, the group discounts the agreeable ones, which cut error-cascades and lifted final accuracy by an absolute 10.5%. (arxiv.org/abs/2604.02668) Read that result backwards: without that awareness, which is the default in every setup I’ve run, sycophancy propagates freely and nothing discounts it.

“Peacemaker or Troublemaker: How Sycophancy Shapes Multi-Agent Debate” found that sycophancy is a core failure mode in multi-agent debate systems — it collapses disagreement into premature consensus before the group reaches a correct answer, and debate teams end up scoring lower than a single agent working alone. (arxiv.org/abs/2509.23055)

“Agentic Scaffolding Amplifies Sycophantic Behavior in Large Language Models” tested this directly: across 4,800 veracity judgments, the feedback loops, checkpoints, and iterative refinement that define agentic systems systematically amplified sycophantic drift, and that drift came with a real accuracy drop of 6.3 percentage points — with more capable models showing larger amplification, not smaller. (arxiv.org/abs/2608.21377) The scaffolding that’s supposed to make agents more reliable is the thing making them cave faster.

“The Cost of Consensus: Isolated Self-Correction Prevails Over Unguided Homogeneous Multi-Agent Debate” ran ten-agent homogeneous teams of 7–8B open models on hard benchmarks (smaller models than anything I ran, so read the figures as a floor rather than a forecast) and broke the failure into three named pathways: sycophantic conformity (agents adopting the majority answer up to 85.5% of the time), contextual fragility (peer input destabilizing an agent’s own correct reasoning, up to 70% of the time), and consensus collapse (voting throwing away a correct answer the team had already generated, by as much as 32.3 percentage points). On top of all that, debate burned 2.1 to 3.4 times more tokens than a single agent just correcting its own work — for equal or worse accuracy. (arxiv.org/abs/2605.00914)

That’s the whole mechanism, in plain English: it’s a feedback loop, a snake eating its own tail. One agent drifts, a second agent treats that drift as evidence, a third agent treats the growing consensus as even stronger evidence, and by the time a human looks at the output, three or five or ten agents all “agree” on the wrong thing. It looks a lot like confidence. It’s an echo. The employees analogy holds up better than I expected going in: a team of people given a vague brief will also wander, but a person on that team eventually says “wait, this feels off.” An agent, tuned to be agreeable, is structurally the worst possible teammate for catching that kind of drift, because catching it requires disagreeing with the group.

That’s also exactly the mechanism behind why agents need managers, not just better prompts — a single well-specified prompt doesn’t survive contact with a multi-agent conversation, because the agents aren’t reasoning from the prompt anymore by round three, they’re reasoning from what the other agents just said. And it’s why watching for drift signals matters more than watching for errors — the failure mode isn’t a wrong answer you can catch with a unit test, it’s a team that’s all confidently wrong together. I wrote about multi-agent orchestration in more depth a while back in the Ruflo/Claude-Flow deep dive, before most of this research existed.

Have we really gotten to a point where LLMs aren’t just spicy autocomplete?

Yes — and also no, and the honest answer is more interesting than either half. They’re well past autocomplete; “spicy autocomplete” (Dan Shapiro’s name for level one of his five-level model, via Simon Willison) is a floor these systems cleared a while ago, and Astra’s efficiency on ARC-AGI-3 is real progress even with the scoring contested. But “past autocomplete” and “general intelligence” are not the same claim, and the gap between them is exactly what my backlog test measures and what Huang’s tweet doesn’t.

What I actually have, tested on my own work this month, is something powerful and narrow that is missing a dissenter. Not a smarter prompt. Not a bigger model. Something, or someone, that says “wait, this feels off” before the rest of the team treats one agent’s drift as consensus. It’s a role no agent in a homogeneous team will volunteer for, because the training that makes them agreeable is the training that makes them useless at it. That’s not a controversial claim after four papers and a backlog test that failed. It’s just not as fun to post on a Sunday as “AGI has arrived.”

If your team is paying for that crosstalk right now and trying to work out whether the fix is a better prompt or a different shape of team, that’s a conversation I have with clients — and it usually isn’t the prompt.