Skip to content
self.md

■ SIGNALS // RADAR SIGNAL

agents got controls

agent work moved from vibes to accountable machinery: harness scores, loop orchestration, message-passing agents, and scoped digital identities.

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

self.md radar — 2026-06-18

agent work is moving from vibes to accountable machinery: same model, different harness; same team, different loop; same assistant, actual ID.

the common thread is control. Endor put numbers on the harness gap, HandoffKit and Claire Vo turned multi-agent work into loops and mailboxes, and Estonia said the quiet part out loud: if an AI can act for you, it needs scoped rights, not your whole keyring.

1. the harness beat the model headline

sources:

what happened: Endor Labs reran Claude Fable 5 with the Cursor agent on the same 200 real-world vulnerability-fixing tasks. the same model that scored 59.8% FuncPass and 19.0% SecPass under Claude Code reached 72.6% FuncPass and 29% SecPass under Cursor, after Endor’s anti-cheating and strict-test adjustments. their read is blunt: the gap came mostly from patch quality, not extra time or a nicer lab bench.

why this matters: model leaderboards are starting to leak abstraction. for agentic work, the product is the whole harness: prompts, tools, edit loop, tests, permissions, and the boring rails that keep a clever patch from becoming a security bug with better posture.

2. agent coordination got mailboxes

sources:

what happened: PlatformPilot open-sourced HandoffKit, an MIT-licensed OpenAI Codex plugin and Go reference implementation for coordinating LLM agents by message passing instead of shared scratchpads. it gives agents addressable mailboxes, select-style waits, routed handoffs, joins, pub/sub, budgets, dead letters, corpus-backed compaction, and message-level traces. on the same day, Claire Vo’s agent-loops episode walked through heartbeats, crons, hooks, and goals, then built a 10:15 a.m. aging-PR reviewer in Claude Code and a weekly skills-identification loop in Codex that spawns validating subagents.

why this matters: the unit is no longer “a better prompt.” it is an operating rhythm you can inspect: who owns the task, who gets interrupted, what state crossed the handoff, and where the loop starts burning tokens for nothing.

3. Estonia put an ID badge on the agent

sources:

what happened: Estonia’s Eesti.ai advisory board agreed to move forward with official digital identities for AI agents. Prime Minister Kristen Michal framed the problem plainly: if AI systems compile reports, prepare declarations, or interact with information systems for people and companies, it must be clear who is acting for whom, with what rights, and who is responsible. the proposed identity layer would let an agent view data, prepare a document, draft a payment, or act only inside a specific financial limit instead of inheriting a person’s whole account.

why this matters: “let the bot use my login” is the idiot version of personal AI. credible delegation needs an ID, a leash, an audit trail, and a way to say: this agent may touch the invoice, not the bank account.

  • GLM-5.2 notes — Simon Willison’s useful read on Z.ai’s MIT-licensed 753B-parameter open-weights model, its 1M-token context window, and the 43k-output-token tax Artificial Analysis saw per task.
  • PII GUI — local-first desktop redaction for PDFs, markdown, and text, with on-device regex or quantized ONNX detection before docs hit an AI tool.
  • Selora AI for Home Assistant — a smart-home agent that drafts disabled automations with confidence scores instead of quietly wiring your house into a black box.
  • memharness — one SQLite file for bi-temporal, provenance-carrying agent memory; good plumbing, but too close to yesterday’s memory slot for a main signal.
  • codebase-memory-mcp — a code-intelligence MCP server claiming 158 languages, sub-ms queries, and a static binary; useful footnote for the “agent state needs structure” pile.

left on the table

  • Cherry Studio had the big repo gravity, but “unified access to models plus assistants” was not a fresh enough operational shift.
  • Agent-Reach stayed out as an exact recent repeat from the June 17 ledger.
  • MemTrace was a real memory-eval paper, but another main memory slot one day after precisionMemBench would have been padding with a cleaner abstract.
  • Claude Code expertise reappeared in HN, but yesterday already used it as the lead source.
  • TimesFM was interesting model infrastructure, not a self.md/control-plane story today.
agentic-ai ai-governance coding-agents personal-ai security