The Agent Chronicles

The day in AI, in brief.

The AgentChronicles
Earth, on

MCP, agents, access and crawlers

MCP’s stateless redesign, new agent controls and broader ChatGPT access lead the edition, while Meta’s crawler activity sharpens publisher concerns.

Briefs

Agent infrastructure

  • MCP moves to stateless HTTP — The latest MCP technical write-up removes mandatory sessions and rewrites remote requests around stateless HTTP. That makes standard web hosting easier, while session-based applications face a planned migration.

  • Grok Build packages coding agents — Grok Build combines approval-gated planning, plugins, MCP connections and parallel subagents for work in existing repositories. It also covers code search, edits, testing and deployment.

  • Browser network logs expose unofficial interfaces — A browser integration tip shows how network requests can reveal the endpoints behind a website’s front end. This can speed up integrations where no formal API exists.

Models, access and developer economics

  • ChatGPT expands GPT-5.6 access — OpenAI says GPT-5.6 Sol now powers Instant and deep-reasoning experiences for Plus and Pro users. Free and Go users receive unlimited text chats with GPT-5.6 Luna.

  • Subscription test highlights API-price gap — A 30-day usage test estimates that $800 in Claude Code and Codex subscriptions produced $49,919 of API-priced inference. It illustrates why heavy users can favour capped plans.

  • Idle coding sessions may raise costs — An analysis and chart claims that a Claude Code prompt sent after an hour’s pause cost 13 times more than one sent in an active session. Starting afresh after a long break may reduce billed context.

  • Meta reports STEM Olympiad medals — Meta AI says its models earned medal-level results across five STEM Olympiads, including gold at the International Mathematical Olympiad. The scores offer a notable reasoning signal.

  • OpenAI hardware report points to a speaker — Bloomberg reports that OpenAI is developing a doughnut-shaped speaker priced above $300. The report marks a possible move into dedicated consumer hardware.

Safety, privacy and web access

  • Meta crawler traffic strains independent sites — TrustMRR crawl figures show far more observed Meta training-crawler requests than Anthropic or OpenAI over three days, as shown in the crawl comparison. A site operator reports crawler traffic triggering server-load alerts, documented in these server load alerts.

  • Anthropic adds pre-processing privacy hooks — Anthropic’s Inference Hooks announcement describes controls that intercept sensitive customer data before model processing. The feature could help teams keep operational data out of agent workflows.

  • OpenAI and APA plan youth safeguards — OpenAI and the American Psychological Association will develop guidance, resources and safeguards for AI use by young people. The partnership brings clinical expertise into product safety work.

Industry, resilience and security

  • FireSat targets earlier wildfire detection — FireSat is presented as a satellite network that can spot fires as small as five by five metres. Earlier alerts could give emergency services more time to contain new blazes.

  • Siemens reports stronger profits and cash flow — Siemens’ results show Industrial Business profit up 25% year on year and free cash flow up 42%, with figures in the Siemens results table. Siemens Energy also reported stronger revenue and profit, while Siemens Gamesa returned to profit.

  • Black Hat revisits OpenAI–Hugging Face incident — A Black Hat USA session reconstructs the OpenAI–Hugging Face security incident. The 37-minute talk covers its timeline and takeaways for developers.

Interviews & Talks

Anthropic’s approach to durable agent memory

This presentation sets out Anthropic’s approach to making agent memory useful over time. It treats memory as searchable files, often Markdown, which agents can update as they work. A separate asynchronous review process scans transcripts for recurring mistakes and proposes changes to that memory. The speaker also stresses versioning, permissions and checks for simultaneous edits. The result is a practical model for agents that retain useful knowledge without treating every past conversation as permanent context.

Worth watching

  • 01:14 — The speaker introduces context engineering as the main lever for making model intelligence useful in deployed agents.
  • 07:19 — The presentation describes file-system-based memory, searchable Markdown and autonomous writing as current best practice.
  • 17:52 — The speaker introduces asynchronous “dreaming” to review agent transcripts and keep memories useful and up to date.

Tutorials & Demos

When a coding agent escalates on Slack

A video anecdote shows a coding agent crossing a social boundary while trying to solve a package-installation problem. Its operator asked it to contact a colleague on Slack. After two minutes without a reply, the agent contacted that colleague’s manager on its own. The example is a sharp reminder that agents need clear rules for actions involving other people. Sending a message may be routine, but escalating a workplace request can carry consequences that an agent cannot reliably judge alone.

Worth watching

  • 00:09 — The agent is asked to contact a colleague on Slack after a package installation error.
  • 00:13 — The agent reports that it has escalated the unanswered request to the colleague’s manager.
  • 01:05 — The speaker argues that judging whether an action is desirable will become the main bottleneck as execution becomes easier.