The Brief
New briefs daily around 7 AM Eastern
Enterprise agent teams report the correction loop dwarfing the agent loop
Enterprise teams presenting at AI Engineer converged on the same shape: the agent loop is the small part of the system. Maersk runs a procedure corpus about twenty times the size of its runtime and earned accuracy through more than 100,000 corrections. Navan dropped output assertions for trajectory scoring once agent reasoning made logs unreadable, and Ironclad discarded token cost as a metric in favor of complexity-weighted merged PRs that clear review and reach a customer. Budget an agent project by the correction and measurement apparatus around the model call, not by the call itself.
- 01Watch
Maersk says 100,000 corrections made its agents accurate
Accuracy came from more than 100,000 corrections rather than a better loop. The ratio reframes staffing on enterprise agent projects: the durable work is curating procedures and corrections, and that work does not shrink when the underlying model improves.
- 02Watch
Navan replaces output checks with agent trajectory scoring
Delayed agent actions also broke the user versus service-principal identity model, so Navan added per-tool-call guardrails and hooks that emit goal and confidence. Output assertions miss the failures that only show up in how the agent arrived at an answer.
- 03Watch
Ironclad measures agent value in trusted merged PRs
Counting token spend without a value denominator rewards spending more. Ironclad's trusted throughput counts PRs that clear checks, review and customer contact, paired with killing flaky tests and capping agent retries so the denominator stays honest.
- 04Read
Tencent releases Hy4 Preview with a 1M-token context
770B total parameters with 49B active, a 1M-token context and 1.56TB of files, a sharp jump from July's Hy3. The chat template accepts only high or no_think for reasoning effort, so cost per call has two settings rather than a dial.
- 05Read
Cursor cuts OpenAI models to 5% of traffic
Cursor puts OpenAI at 5% of its traffic and is leaning on Grok 4.6 and Claude instead. GPT-based workflows inside Cursor need a migration plan, and the cutoff sets a precedent for model vendors withdrawing access from coding tools they compete with.
- 06Read
OpenAI chip reportedly beats Nvidia Rubin on power efficiency
SemiAnalysis reports the in-house accelerator taped out in nine months and leads Nvidia Rubin on tokens per megawatt. If the number holds through volume production, inference pricing and CUDA lock-in both move over the next capacity cycle.
- 07Read
Claude Code limits settle 17% below the temporary boost
The increase covers Pro, Max, Team and seat-based Enterprise, but it settles 17% below the temporary boost running now. Agent workloads sized against this week's headroom will not fit once the standard limits take over.
A dated brief from the vibe-coding frontier. Today’s Intel.