Skip to content
self.md

■ SIGNALS // RADAR SIGNAL

agents got carded and graded

Claude identity checks, Lighthouse agentic browsing audits, and ANMA module contracts all point at the same shift: agent control is moving into enforceable surfaces.

Published Updated Read 4 min
■ [!] ON THIS PAGE ▼

self.md radar — 2026-06-22

the control surface moved closer to the thing doing the work: Claude account access, Chrome’s agent-readable page checks, and repo boundaries that can actually trip an agent before it wanders off.

Anthropic is putting identity checks into some Claude flows. Chrome is teaching Lighthouse to inspect whether a page is usable by agents, not just humans. ANMA is the repo-level version of the same mood: stop trusting the prompt, compile the boundary.

1. Claude access got an ID check

sources:

what happened: Anthropic’s updated Claude Help Center page says identity verification is rolling out for “a few use cases.” the prompt can appear when a user accesses certain capabilities, during routine platform-integrity checks, or for other safety and compliance measures.

The verification partner is Persona Identities. the checklist is not subtle: a physical government-issued photo ID, a phone or computer with a camera, and possibly a live selfie; Anthropic says the process usually takes under five minutes and is used to prevent abuse, enforce usage policies, and meet legal obligations.

why this matters: agent access is no longer just a token, a seat, or a rate limit. for some capabilities, the account now has to resolve back to a verified human body, which makes “personal AI” less anonymous right where the tools get more capable.

2. Lighthouse started grading pages for agents

sources:

what happened: Chrome’s Lighthouse docs now include an experimental Agentic Browsing category. it evaluates whether a site is built for machine interaction through deterministic audits, with checks around WebMCP integration, registered WebMCP tools, schema validity, llms.txt, accessibility for agents, and layout stability.

The category does not use the usual weighted 0–100 Lighthouse score. Chrome says the standards are still emerging, so the report shows pass/fail signals and a fractional readiness view instead; WebMCP audits require Chrome 150 or later plus the WebMCP origin trial.

why this matters: a web page becoming “agent-ready” is turning into something you can run in a boring audit panel. that is good news for operators, and bad news for the old habit of pretending a chatbot can safely click anything just because the DOM exists.

3. module boundaries got teeth

sources:

what happened: ANMA turns plain-YAML module contracts into generated CLAUDE.md, hooks, and CI checks for coding agents. the pitch is narrow enough to be useful: declare which modules may depend on which public interfaces, then make the same rule visible to the agent and enforceable by tooling.

The benchmark detail is the signal. on a Python payments-boundary task, Claude Haiku 4.5 violated the boundary in 13 of 19 scored control runs; with ANMA installed, it violated it 0 of 20 times. the follow-up numbers for TypeScript and Go are more careful, but still point at the same operator lesson: cheaper agents need fences more than encouragement.

why this matters: repo instructions are easy to admire and easy to ignore. a boundary that can block a tool call, fail CI, and survive drift is a different object: less pep talk, more guardrail with a bruise pattern.

  • DeerFlow — ByteDance’s 2.0 super-agent harness bundles subagents, memory, sandboxes, skills, MCP integration, and a message gateway; too broad for the lead slot, useful as a map of where agent runtimes are heading.
  • FreeLLMAPI — one OpenAI-compatible /v1 endpoint across 16 free provider tiers, with routing, failover, encrypted keys, and per-key usage tracking; the cheap-token hustle is becoming infrastructure.
  • Patterns for Building Cybersecurity Evals — Eugene Yan’s eval recipe keeps the shape clean: sandboxed target, difficulty-controlling inputs, tools, and a grader.

left on the table

  • DeerFlow 2.0 RC stayed supporting because this week already spent plenty of oxygen on harnesses, mailboxes, and long-running agent work.
  • BrowserClaw was too close to yesterday’s WebMCP-in-the-page story, and the repo read more like a wrapper than a new control surface.
  • Apertus is interesting sovereign-model infrastructure, but the main launch is from 2025, not today’s shift.
  • Ask HN: do you give AI coding agents their own GitHub account? matched the identity theme, but a small HN thread was weaker than Anthropic’s own verification page.
agent-governance browser-agents coding-agents personal-ai