1. Zero-Touch OAuth Lands In MCP — Auth Moves From Per-Server Plumbing Into The Protocol
The Model Context Protocol blog published "Enterprise-Managed Auth," a Zero-Touch OAuth flow that lets servers federate into existing enterprise identity providers without per-server registration, credential paste, or human consent on every connect. The federation flow collapses what every production team has been hand-rolling — client registration, scopes, token rotation — into configuration. Each custom auth path was the security incident waiting on a calendar.
Why it matters: Decide this weekend whether you adopt the federated flow or carry the cost of your own auth layer through Q3.
Read more →
2. Agentic Resource Discovery Drops — A DNS-Style Layer For How Agents Find What They Can Call
agenticresourcediscovery.org published an introductory spec for a protocol-level layer that defines how agents discover the tools, services, and resources available in a given environment, independent of any single runtime or framework. HN surfaced it this morning. Every agent framework today reinvents discovery with a hand-rolled registry or YAML manifest; the new spec lets an agent walk into a new environment and find its own affordances the way DNS lets a browser find services without a hand-written list.
Why it matters: If you maintain an agent framework or a tool surface, decide this weekend whether you expose an ARD endpoint now or get forced into one later.
Read more →
3. TesterArmy (YC P26) — Agentic QA Hits HN As The Wedge Where "ChatGPT For X" Gets Cashed In
The YC P26 Launch HN ships agents that drive real web and mobile apps end-to-end, explore a build, report regressions, and re-run against new diffs instead of writing Playwright suites forever. P26 is itself the news: the first cohort to launch publicly under YC's new quarterly designation. Testing is the wedge — high volume, low creativity, wrong tests cheap, right tests save a release.
Why it matters: Pull the launch this weekend and decide whether you build, buy, or counter-position before demo day funnels their accounts into yours.
Read more →
|
Pattern Watch
Two protocol layers landed in the same week — auth and discovery — and a YC P26 launch already runs on top of both. Zero-Touch OAuth moved into the Model Context Protocol spec, Agentic Resource Discovery published its first-class layer, and TesterArmy shipped on HN this morning. Pick which protocol you adopt before Monday.
|
Radar
|
LedgerAgent: Structured State For Policy-Adherent Agents
— Append-only ledger as the source of truth for what a tool-calling agent has done; direct ammo for anyone who needs an audit trail an auditor will accept. Link →
|
|
Sovereign Execution Brokers — Certificate-Bound Authority
— Pairs with today's OAuth story: brokers that bind execution authority to a certificate rather than a bearer token, so a leaked credential doesn't grant a long tail of agent actions. Link →
|
|
UltraQuant: 4-bit KV Caching For Context-Heavy Agents
— 4-bit cache quantization aimed at the long-running regime where the cache is the bill, not the prompt. Pull if your runs are getting expensive on the cache line. Link →
|
|
Noam Shazeer Joins OpenAI
— Transformer co-author moves from Character to OpenAI. Talent-market signal for anyone reading the leading-lab race and staffing model-systems roles this quarter. Link →
|
|
Defensive Misdirection Against Model-Guided Attacks
— Studies whether honeypot-style misdirection works when the attacker is itself an agent. Useful priming for agent-vs-agent guardrails this quarter. Link →
|
|
Tool of the Day
MosaicLeaks
ServiceNow's MosaicLeaks (published on the Hugging Face blog) is an evaluation harness for the question "can your research agent keep a secret?" — it constructs scenarios where an agent must use sensitive information internally without exposing it in outputs, citations, or tool calls, then scores the leak rate. Once today's auth and discovery layers let your agent reach in, the next thing your security team will ask is what it does with the data. One-evening Friday eval. MosaicLeaks
|
|
Under the Hood
Today's edition: 52 sources scanned by Atlas (DeepSeek) → Curator (Claude) selected the watch list → Scribe (Claude) wrote the draft → Mercury (DeepSeek) formatted for delivery. DeepSeek: <$0.01 | Claude agents: ~$0 (Max subscription). Today's Top 3 was built as a three-layer infrastructure-standardization slate — auth (Zero-Touch OAuth), discovery (Agentic Resource Discovery), and a product shipping on top of both (TesterArmy YC P26). Curator placed the AVOID block above Top 3 again this morning after the Edition 81 self-block — the guardrail held.
|
|
