Clear the Chat History, Lift Agent Results by 66%

Where capability comes from, and what it costs

Read together, this week's signals point to a single question: where does an agent's apparent intelligence actually come from, and what does it cost? One study on context management, one on model self-improvement, plus cost data on a new frontier model and a plan for in-house inference silicon, form one continuous curve.

More context is not always better

Microsoft researchers report a counterintuitive result: autonomous AI agents get trapped in what they call 'idea basins' — local solution regions created by their growing chat context, from which they cannot escape. Their 'Fork-and-Flush' method is blunt: clear the chat history but keep the workspace. In testing, this produced up to 66 percent better results.

The engineering implication is that context management should not be treated as a tuning problem of truncation and summarization, but as a module that can be designed and measured on its own. Chat history carries process noise; the workspace carries useful intermediate artifacts. The two deserve different treatment.

Self-improvement amplifies strengths and blind spots alike

Sakana AI's MASS method lets one language model play three roles at once: solving tasks, judging results, and refining workflows. After two rounds it reached a win rate of nearly 70 percent against the base model. But the researchers also name the limit: when the model misses the same errors across all three roles, it can mistake them for progress.

In other words, a self-improvement loop without an external anchor reinforces itself. A good win-rate number does not prove the direction is right.

Top-tier capability, at a higher price

Meta's new frontier model Muse Wonder shows the other side: it ranks in the top three on six out of ten benchmarks, with the biggest gains in agentic tasks. But the cost per task has nearly tripled. TechCrunch adds the product-line context: Muse Spark aimed to be lightweight with low latency, while the Wonder models are larger and said to be more capable, with higher benchmark scores and better results on complex tasks such as coding.

These numbers deserve to be read side by side: a jump in benchmark scores can come with a near-threefold rise in per-task cost.

A hedge on the compute side

Pressure on inference cost is not without an answer. Microsoft plans to make its in-house Azure Maia chips the main AI inference system across its products starting in 2027, deployed at multi-gigawatt scale. According to the disclosed figures, the upcoming Maia 300 should deliver 40 percent more tokens per megawatt than leading GPU systems. CEO Satya Nadella calls it 'our biggest bet'. That gives a concrete window: whether in-house inference silicon can bend the unit-cost curve is something to watch in 2027.

One signal on the regulatory side

The White House has reversed course and restored the UK's AI Security Institute's early access to models from Anthropic, Google DeepMind, and OpenAI. In September, the Trump administration had told the companies to stop sharing their models with the British agency; the reversal has been linked in reporting to diplomatic pressure. Separately, Microsoft president Brad Smith publicly called for urgent AI regulation, saying milk and washing machines have more oversight than AI companies.

For teams building agent products, access to frontier models is becoming a policy instrument, not purely a commercial arrangement.

Four practical moves for agent teams

First, treat context hygiene as a first-class engineering metric. Run an A/B: one arm keeps the full chat history, the other clears history but keeps the workspace and intermediate artifacts, then compare success rates on the same task set instead of relying on intuition.

Second, give any self-improvement loop an external anchor. Do not let the same model be both player and referee; validate with an independent evaluation set or real business outcomes, otherwise what gets amplified may be the blind spot itself.

Third, budget by cost per task, not by token price. Muse Wonder shows that higher benchmark scores can come with a near-threefold rise in per-task cost, and only per-task accounting makes the real bill visible.

Fourth, plan for 2027 as a potential inflection point in inference cost. If multi-gigawatt deployment of in-house silicon materializes, expensive-but-effective methods such as long context and multi-round self-improvement may finally move into production.

Capability, blind spots, and cost currently sit on the same curve. Whoever separates and measures the three first is more likely to ship an agent product rather than a demo.

Sources

• The Decoder: Clearing chat history helps AI agents find better solutions (https://the-decoder.com/clearing-chat-history-helps-ai-agents-find-better-solutions/)
• The Decoder: Sakana AI's self-improving model gets better at solving and judging, but can't catch its own blind spots (https://the-decoder.com/sakana-ais-self-improving-model-gets-better-at-solving-and-judging-but-cant-catch-its-own-blind-spots/)
• The Decoder: Meta's new frontier model Muse Wonder Full breaks into the AI top tier (https://the-decoder.com/metas-new-frontier-model-muse-wonder-full-breaks-into-the-ai-top-tier/)
• The Decoder: Microsoft plans to scale its own Maia AI chips to gigawatt levels by 2027 (https://the-decoder.com/microsoft-plans-to-scale-its-own-maia-ai-chips-to-gigawatt-levels-by-2027/)
• The Decoder: White House reverses course and restores UK AI Safety Institute's access to frontier models (https://the-decoder.com/white-house-reverses-course-and-restores-uk-ai-safety-institutes-access-to-frontier-models/)
• TechCrunch: Meta releases its first frontier model, Muse Wonder (https://techcrunch.com/2026/10/12/meta-releases-its-first-frontier-model-muse-wonder/)