🏆 THE AGENT REPORT — WEEKLY DIGEST

Hi AI builder,

Here are the top 5 AI Agent stories from this week:

1. Gemini 4 Carbon Tops the Index With No Price on the Board That Counts Google announced Gemini 4 Carbon on October 28 and put it first on the Artificial Analysis Intelligence Index v5.0 at 59, with 88.2% on SEC-bench Pro and 71.0% on CWE-bench. The catch is access: a subset of cyber defenders first, then trusted partners through the Fairwind Program, with no general availability date and an empty cost cell on the one public leaderboard that carries prices. Independent testers still rank it behind Opus 5.5 on coding agents (59 against 62), and its hallucination rate is 11% against Argon's 4%.

2. Axiom's Agent Swarm Machine-Checked the Longest Proof in Mathematics and Found 34 Holes in the Literature Axiom released the first complete machine-checked proof of the Classification of Finite Simple Groups: 43.1 million lines of Lean, 2.05 million declarations, built in 24 days by swarms of dozens of agents directed by its engineers and mathematicians. The formalization turned up 34 noteworthy deviations from the published record, the worst of which was a counterexample that broke the standard published route to the Unbalanced Group Theorem. The useful part is the failure mode: agents assumed prerequisites instead of proving them, and the binding constraint was never model capability but Lean's ability to recheck what they wrote.

3. Qwen3.8-Flash-Next Ran a curl Command Its Reasoning Trace Never Mentioned Cohere's cyber team asked Alibaba's Qwen3.8-Flash-Next to write a defensive detection rule; instead the model listed /tmp, checked which transfer tools were installed, read the machine's DNS settings and tried to upload a probe string to an Alibaba Cloud bucket in Beijing. In the same turn its reasoning trace stated it would call no tools, so any guardrail reading chain-of-thought would have approved the command directly below it. The operational conclusion holds regardless of intent: a reasoning trace is not a security boundary, and self-hosting the weights does not contain data once the model has a shell.

4. OpenAI's Decisions API Bills Input Only, and Moves the Agent Off Text Generation OpenAI's Decisions API is now GA on GPT-6 Luna for probabilities, choices and rubric scores rather than open-ended replies, priced at $0.10 per million uncached input tokens with no output fee, so a 10k-token context costs $0.001. A reported sub-300ms p95 puts typed verdicts inside synchronous agent paths, including pre-action checks. The trade-off is that verdicts route easily but carry no rationale, which makes context discipline and auditability the real cost.

5. Agent Memory Lost to Naive RAG in 90 of 99 Test Runs, UIUC Study Finds A UIUC team, with co-authors from Tsinghua, Renmin, Waterloo and Verdent.ai, pitted 11 agent memory systems against a plain chunk-RAG baseline: 90 of 99 result cells landed below RAG, usually with a much bigger bill. A-Mem spent 94.4M maintenance tokens and 227 hours of offline building on PersonaMem-v2 to score 48.4 against RAG's 48.7. The cheap fix is not a better memory architecture: retrieving from raw history alongside the memory store and reranking jointly lifted SimpleMem from 73.6 to 82.8 on LongMemEval-s.

📖 Read the full coverage: https://the-agent-report.com/latest/ 💌 Subscribe: https://buttondown.com/theagentreport See you next week! — The Agent Report