The Product Desk

Why the group fails, mechanically

A staff engineer opened the eval log expecting thirty different approaches to one coding task. She read the branch names first, because that is the fastest tell. LLM outputs are far less varied than orchestration diagrams assume: hand 30 agents the same coding task and 18 of them name their git branch identically. Cloning one mind four times does not produce four perspectives. It produces one perspective with four votes. Aggregation subtracts on top of that. Blending several models' answers preserved only about a quarter of the good ideas a single model had already generated. Averaging pulls toward the answer the shared evidence supports, and in a hidden-profile task that is precisely the wrong answer.

Azeem Azhar's diagnosis separates the thing being pitched from the thing being built. The missing ingredient is social, not technical. Human groups survive this failure mode because they have reputation, recourse, and protection for the lone dissenter. An agent holding the decisive fact has no standing, no track record of being right, and no path to escalate over apparent consensus. Azhar and his co-author concede the coordination problem may be fixable, but say plainly that no fix is yet clear.

The outlier that changes the shopping list

One model family, Mythos 5, scored roughly 85% on the identical task while the others sat in the 17-36% band, and the write-up states that nobody knows why. Read that as a procurement instrument, not a design principle: decision quality on this task class moves by tens of points on model choice alone. The forcing function is a line item in the next model evaluation asking for a hidden-profile result, plus a standing rule against hard-coupling a shipped product to a mechanism no one can explain.

The absorption clock sitting on top of it

Latent.Space's reporting gives a second reason to keep this work thin and swappable. Tool calling and context compaction have already migrated out of harness code and into model weights. Multi-agent orchestration, tool selection and memory are flagged as the next candidates. The epic with the weakest evidence behind it is also the capability most likely to arrive free in a checkpoint. That is the worst available place to spend a quarter of engineering capacity.

The lane nobody has taken

The gap the researchers describe is buildable, and it is not a model problem. Trust weighting per agent, evidence provenance, and a dissent-escalation path that routes a minority-held fact to a human or a full-context arbiter is a product, with a ready-made narrative and no incumbent. It is also the one part of orchestration the absorption schedule does not threaten, because it governs who is accountable rather than what the model can compute. Two axes for the sprint decision: does the work add agents, or add accountability between them, and would a better checkpoint make it redundant. One cell survives both questions.

Multi-agent is not an architecture upgrade until someone builds reputation and dissent into the orchestration layer. Until then it is an accuracy tax on decisions one agent with the full file gets right.

Carry the caveat into the exec conversation. The 17-36% and ~85% figures arrive without linked methodology, in a truncated preview. Run them as a hypothesis tested in-house this sprint, not as a citation to lean on in a review.

Both generations at once

A capacity planner reopened the 2027 rack model this week and found the escape hatch gone. The increase covers GB300 and VR200, current generation and next, both inside the same 2027 delivery window. Repricing one node is routine. Repricing the whole roadmap removes the "wait for the next platform" line infrastructure plans quietly assume, and it says Nvidia sees little substitution pressure at rack scale. Server makers are the transmission channel, so the squeeze lands in OEM and customer P&Ls before it touches Nvidia's margin line.

The GPU cloud market has already split on the bet. Per Catherine Perloff's companion reporting for The Information, Nebius and CoreWeave are touting short-term deals while AWS pushes long-duration commitments. Those are opposite forecasts of where prices go next. Picking one sets the cost basis under every AI feature shipped in 2027.

Depreciation ambiguity is the buyer's leverage

Nvidia is sending mixed messages on chip useful life in the same quarter it raises prices. A higher asset cost amortized over an unknown life is the worst input a TCO model can take, which makes it the strongest item on the buyer's side of the table. Contractual useful-life assumptions, price caps and generation-migration rights are all askable this cycle. A vendor who refuses every one of them has answered the question.

The ceiling arrived in the same quarter

THE DECODER's Frontier Radar supplies the other jaw. Anthropic charges roughly 4.4x the market average per token while Kimi K3 and GLM-5.3 have closed to striking distance of the best US models; GLM-5.3 tops some rankings despite a delayed release, and Deepseek's experimental Flash vision model rivals Opus 4.8 on agent benchmarks. Jurisdiction keeps some of those models out of production stacks. It does not keep them out of procurement, where they set the price a buyer now believes is achievable. Anthropic reportedly passed OpenAI at $65B+ annualized revenue, and reporting that its bioweapons filter sat inoperative for nearly a year removes part of the trust premium that justified the multiple.

Software is the only deflation left

Volume will not rescue the model. Exponential View measured token price elasticity at 1.2-1.8: a 10% price cut lifts usage 12-18%, enough to grow spend, nowhere near a Jevons explosion. What gets pitched as inference cost reduction is mostly scheduling. SGLang's prefix-aware scheduler reuses every shared prompt prefix (system prompt, tool definitions, conversation history) through a RadixAttention tree instead of recomputing identical tokens each turn. Agent loops overlap by construction, so an engine without prefix reuse pays full compute for the same tokens every turn. ByteByteGo's breakdown puts Ollama's FIFO queue at the other extreme: correct on a laptop, structurally wrong for multi-tenancy.

Sources agree on direction and disagree on precision. Chris Short's roundup tracks AI compute financing escalating from $35 billion for one gigawatt in June to a sought $60 billion-plus SPV in August, with Broadcom backstopping senior tranches. Cheap inference is currently a credit-market product. The Information's own contrarian read notes the increase rests on two anonymous sources and applies to some flagship systems. The forcing function is a plan that survives 8% as well as 17%: delete the free-deflation cell from the model and make cost per successful task an owned metric beside latency and quality.

Compute is not getting cheaper on its own any more. Every remaining point of AI margin has to be engineered out or negotiated down.

Capability arrives without a changelog

The morning after a frontier checkpoint drops, someone on the platform team re-runs the bench and deletes a block of prompt scaffolding that stopped earning its keep. No release note told her to do it. Lukasz Kaiser, a co-inventor of the Transformer, described last winter's step change on the Unsupervised Learning podcast this way: "the harness changed and a little post-training changed and then new pre-trained models came... but it felt like a big jump which is not that easy to pin down what did it." If the co-inventor of the architecture cannot attribute the gain, no vendor roadmap will do better. The forcing function is a 72-hour clock: after any frontier checkpoint, re-run the bench, then answer two questions in writing. Which scaffolding is now redundant, and which autonomy increment just became viable.

Separate what is being pitched from what is being done. The absorption is documented. OpenAI's codex-1 was "trained using reinforcement learning on real-world coding tasks in a variety of environments," and GPT-5.1-Codex-Max launched as the first model "natively trained to operate across multiple context windows through compaction." Meta's 2023 Toolformer thesis — tool use trained in rather than prompted from outside — is now shipping as product. Train, absorb, shed, repeat, on the labs' clock rather than the planning cycle.

Sort the backlog into three buckets

Why harness work still pays this quarter

Harness engineering is not worthless. It is perishable, which is a different planning problem and a much better one to name out loud. OpenAI tripled GPT-5.6 Sol's ARC-AGI-3 score from 13.3% to 38.3% using only retained reasoning and compaction, with no new weights. The counterexample is about timing, not ambition. Devin's first version tested at roughly 15% success in Answer.AI's evaluation. Claude Code shipped a near-identical autonomy thesis in February 2025, after the reasoning-model crossover, and reached about $1B ARR within six months. Same product idea, different release date. So write autonomy into the PRD as an explicit gate: maximum task horizon equals the step count at which measured per-step reliability still clears the target end-to-end success rate. At 95% per step, 20 steps compounds to roughly 36%.

The same pattern one layer down

Docker proved its own hypervisor first in Docker Sandboxes for AI agents; the beta shipped in Desktop v4.86 with Linux GA targeted for October 2026. After GA, Docker is the default substrate for agent sandboxing and the negotiating position thins considerably. Encore's parallel microVM effort shows the do-it-yourself path works and also marks its ceiling. Firecracker requires /dev/kvm, which macOS does not have, and Apple gates VM snapshotting behind a private entitlement while its own validator falsely reports support. Any roadmap line reading "snapshot and restore the dev environment" needs verifying before anyone commits to it.

Two honest caveats. The forecast that every agentic company ships a user-editable interrupt-and-delegation policy within a year rests on a single author and a handful of early Anthropic features, so size a v0 as discovery, not a platform commitment. And self-improving harnesses trained like models could collapse the harness/model distinction entirely, obsoleting hand-built scaffolding faster than any absorption schedule implies. Both risks argue the same way: thin harness, thick human-facing layer.

A system prompt that grows quarter over quarter while models improve is a team paying twice for capability it already licenses.