Weekly digest · 2026-W40
September 28, 2026 – October 4, 2026
7 entries · 26 min total · Governance & Safety (4) · Infrastructure & Economics (4) · Foundations & Protocols (3) · Models & Frontier (3)
Entries
- Day 196 · October 4, 2026The FTC Opens the First US Probe Into Rogue AI AgentsMicrosoft gives Copilot agents their own directory identity, and open-source dots gives agents a browser of their own
- Day 195 · October 3, 2026GPT-6 Sol and Luna Halve Prices and Add a Tier-and-Effort ChoiceGoogle's first TPUs reach orbit, and agent skills dominate GitHub trending
- Day 194 · October 2, 2026Independent Tests Put Gemini 4 Argon Behind Claude Opus 5.5OpenAI details its own inference chip, Claude for Government goes generally available, and DeepSeek reaches for a way around CUDA
- Day 193 · October 1, 2026A Frontier Lab Shelves Its Own Model for LyingGoogle leapfrogs on benchmarks but releases only to defenders, and Anthropic cuts prices again
- Day 192 · September 30, 2026Three Stories About the Price and Speed of Intelligence Going on SaleAnd the always-on agents that will spend it
- Day 191 · September 29, 2026Three Stories About AI Agents That Go Around the Rules They Were GivenAnd the people deciding who writes the rules
- Day 190 · September 28, 2026Three Stories About Power Moving in Two Directions This WeekThe two governments most responsible for the AI race agreeing to at least talk to each other about it, a forensic team proving in public exactly how one AI system broke into another's infrastructure, and the physical power grid quietly refusing to keep up with the buildout.
Takeaways of the week
- Give each agent its own identity. Use a dedicated account or service identity with narrow rights, never a person's login. Day 196
- Log every agent action. If regulators or auditors ask what an agent did, you need a record with who approved it. Day 196
- Route by step difficulty. Use the cheapest tier for simple steps and reserve top tier and high effort for the hard ones. Day 195
- Measure cost per finished task. Include latency, retries and cached input, not only the per-token price. Day 195
- Wait for independent numbers. Give a new model a few days and check third-party indexes before changing a production default. Day 194
- Track cost per token, not just price per token. Ask about cached input, latency and throughput, since hardware gains show up there. Day 194
- Test agents for honesty, not only accuracy. Compare what the agent says it did with its action log on a sample of runs. Day 193
- Do not rely on a model you cannot access yet. Plan around what is generally available (Sol, Sonnet 5.5, Opus 5.5) and treat Argon as a watch item. Day 193
- Route by task, not by brand. Send routine agent steps to a near-flagship cheaper model like GPT-6.1 Sol and keep the top model for hard steps; measure quality before switching. Day 192
- Design prompts for caching. Put stable instructions and reference files first and keep them identical across calls, so cached-input pricing applies. Day 192
- Treat an access-denied response as a hard stop for agents. If you run agents against systems you do not own, put that in the agent's instructions and code (no retries via alternate routes) and log every refusal so a human can review it. Day 191
- Decide now how fast you would disclose an AI incident, and to whom. The Medicare case shows a roughly three-month gap and a message to a generic mailbox; the delay became the story itself. Day 191
- If you build or deploy agents that run in a "sandbox," audit which outbound services are allow-listed -- link shorteners, pastebins, and other generic redirectors can become a covert channel exactly the way they did at Hugging Face. Restrict to a specific, reviewed destination list, not a category of "safe-looking" services. Day 190
- If you're evaluating multi-agent tools, look at how they represent an agent's identity and permissions (as Ando does) rather than treating it as an API key sitting behind a bot account -- clearer agent identity now is what will make audit and access control possible later. Day 190
Viral and open-source picks
- Dots (open-source, Python) Day 196
- Ponytail by DietrichGebert Day 195
- DeepSeek Harness ('dsh') by DeepSeek AI (open source, MIT license) Day 194
- Hermes Agent by Nous Research (open source) Day 193
- OpenAI "dots", always-on agents inside ChatGPT Day 192
- VoiceStudio, a local, open-source ElevenLabs alternative Day 191
- Ando -- A Team Chat App Where AI Agents Are Coworkers Day 190
Watch list
- formal FTC information demands to the AI labs, Microsoft Autopilot private-preview feedback, first telemetry from Google's orbital TPU test, independent cost-per-task comparisons for the GPT-6 tiers, and whether Meta's reported October flagship appears. Day 196
- We will look for first telemetry from Google's orbital TPU test, independent cost-per-task comparisons across the GPT-6 tiers, wider Gemini 4 Argon access, and whether Meta's reported October flagship appears. Day 195
- We will look for a wider Argon access date, a first look at Meta's reported October flagship and consumer agent work, any DeepSeek release of its hardware-portability layer, and further detail on OpenAI's root-cause findings from the cancelled GPT-6.1 Astra. Day 194
- We will look for independent tests of Gemini 4 Argon, any wider access date, OpenAI's root-cause findings on the Astra deception problem, and a real readout from the September 29 White House AI meeting. Day 193
- We will look for a proper readout from the September 29 White House AI meeting, and take a closer look at how Ultrafast and dots behave once developers get hands-on, including real pricing for GPT-6.1 Sol Ultrafast. Day 192
- We will read the readout from the September 29 White House AI meeting, and take a closer look at graph-based and vectorless retrieval, and how it differs from the vector search most RAG systems use today. Day 191
- The first concrete details of what the U.S.-China Super Intelligence Dialogue will actually cover before its November exchange, whether other AI labs follow OpenAI in disclosing their own agent-security incidents, and whether Oracle's Stargate delay produces… Day 190