self.md radar — 2026-06-09
the assistant story got useful where it had receipts: a cloud boundary, an installable skill catalog, and local tripwires before the agent spends money or leaks a key.
Apple supplied the loud object, but the quieter artifact was the security post. Google shipped agent skills as something you install, not a doc you hope the model remembers. and the small tools on the floor kept saying the unglamorous part out loud: if the agent can act, the boundary has to run before the damage.
1. Apple put Gemini behind a privacy boundary
sources:
what happened: Apple introduced Siri AI for developer testing today, with a user beta planned later this year. the promised surface is big: personal context across Messages, Mail, Photos, and third-party Spotlight integrations; onscreen awareness; web answers; app actions; a dedicated Siri app with conversation history; and writing tools that can draft or edit almost anywhere the user types.
The architecture post is sharper than the product page. Apple says it collaborated with Google to use technologies behind Gemini to build the next generation of Apple Foundation Models, then worked with Google and NVIDIA to extend Private Cloud Compute to Google Cloud systems using NVIDIA GPUs. the demanding bucket is named plainly: agentic tool-use and complex reasoning, with Apple saying PCC on Google Cloud still keeps stateless computation, no privileged runtime access, non-targetability, transparency, public binaries, research tooling, and live research-mode node access.
why this matters: the personal assistant is now a routing problem with a privacy claim attached. if Siri can see the screen, search the personal archive, touch apps, and call cloud models, the trust surface is not the answer bubble; it is the attested path between the device, the model, the tool, and the log.
2. Google made skills a vendor surface
sources:
what happened:
Google’s google/skills repo is now a very visible official shelf for Agent Skills around Google products and technologies. the README says to install it with npx skills add google/skills, then choose specific skills from the repo. the current list includes Gemini API, Gemini Interactions API, Managed Agents API, Skill Registry API, Cloud Run, BigQuery, Firebase, GKE, AlloyDB, Cloud SQL, and Google Cloud Well-Architected Framework skills.
At verification, the repo was Apache-2.0, created March 31, and sitting at 12,612 GitHub stars. its recent commits added and tuned cloud agent-platform skills, including endpoint management, inference, and onboarding work in the last week.
why this matters: skills are turning into a distribution format for operational memory. that is good news for agents that need exact cloud recipes, and bad news for anyone pretending “read the docs” is still enough when the working interface is an installable context pack.
3. agent safety moved into preflight tools
sources:
what happened: Guardian Runtime describes itself as a local-first FinOps and security firewall for AI applications. it runs as a local reverse proxy or SDK, scans outbound LLM payloads before they leave the machine, blocks secrets and PII, enforces hard budgets, tracks cost, and trims payloads. the docs show the intended failure mode clearly: a Claude Code or Aider-style agent points at localhost, hits a daily budget or secret detector, and gets stopped before the provider receives the mess.
Pokayoke attacks the same problem from inside the repo. its docs frame repo conventions as checks that sit between linting and project reachability, with local TypeScript rules and an agent SKILL.md so humans and agents can run, understand, and repair the checks instead of arguing with a paragraph in AGENTS.md.
why this matters: nobody has enough discipline to prompt their way out of every loop, secret, and house-style violation. useful agent control is getting boring on purpose: proxy, budget, scanner, local rule, failing check.
supporting links
- Syll paper — the open-source personal automation paper ties MCP/API tools, CLI execution, web surfaces, and desktop GUI control into one harness with demonstrations compiled into reusable skills.
- OpenLTM — a local Claude Code memory plugin with SQLite, semantic recall, session hooks, decay, and no cloud account; strong support, but too close to the memory run from the last few days.
- AI Boost — an MCP expertise layer for saving and re-injecting conventions; useful as a curated-memory contrast to automatic memory plugins.
- Panniantong Agent-Reach — a CLI for giving agents read/search access across Twitter, Reddit, YouTube, GitHub, Bilibili, and XiaoHongShu without official APIs.
left on the table
- last30days-skill had the right research-agent smell, but it was a repo gift, not a stronger main signal than Apple’s infrastructure move.
- OpenLTM and AI Boost were kept out of the numbered slots because memory/context has already eaten a lot of this week.
- Do AGENTS.md files help coding agents? and the AGENTS.md evolution thread were too close to the recent rules-and-skills coverage.
- Guardian Runtime’s HN post was useful for discovery, but the published signal leans on the repo and docs because those held the inspectable claims.
Related self.md routes
- Personal AI OS tools — the control-plane map for personal agents, receipts, memory, and tools
- AI coding assistants — compare coding workbenches by review surface, permissions, cost, logs, and escape hatches
- Best Claude Code plugins — choose the Claude-specific extensions worth installing, and the ones to skip