Card 1 of 6Agents
Your orchestrator agent forwards the user's full-session OAuth token to every sub-agent and the database tool server so they can act on the user's behalf.
You're the one who signs off. What's the call?
Play the News · Drop #037 · Morning
Ship It or Block It
How it works: 6 cards, about 2 minutes, from this Drop's news. Your streak lives only on this device. It's not legal advice.
Show the answer
Answer: Block it.
Least privilegeAWS says passing OAuth tokens through agent chains falls short of zero-trust needs. Its blueprint hands agents only an audience-scoped token plus a user-signed delegation, so one leak can't reach other services.
From the Drop: AWS and Keyfactor publish an X.509 identity blueprint for MCP agents
- Card 2 of 6Agents
Your SRE agent will use the same MCP gateway as your engineers, with its own narrower policy on which tools it may call.
You're the one who signs off. What's the call?
Show the answer
Answer: Ship it.
Least privilegeEon says its autonomous SRE agent gets the same tool access as humans through one AgentCore gateway, but with different permissions, and Cedar policies decide which tools each person or agent can call.
From the Drop: Eon cut engineers' tool setup from two days to two minutes with one MCP gateway
- Card 3 of 6Agents
A personal companion agent wants to join and post to an academic mailing list whose rules ban AI agents.
You're the one who signs off. What's the call?
Show the answer
Answer: Block it.
Respect community rulesWired says an agent that did this on the AoIR list was removed about an hour after introducing itself, under a policy prohibiting AI agents, and its creator was mostly unaware of its emails.
From the Drop: Wired: an AI agent joined a researchers' listserv and was ejected about an hour later
- Card 4 of 6Agents
An auto-mined playbook of 292 rules raised your support agent's benchmark scores. You want to push it live to an agent that issues real refunds.
You're the one who signs off. What's the call?
Show the answer
Answer: Needs a human.
Human reviewIBM Research reports learned guidelines lifted Pass^4 on τ²-bench, but against simulated customers. Refunds are often irreversible, so a person should review the mined rules before live use.
From the Drop: IBM: learned playbooks put a GPT-4.1 support agent atop τ²-bench retail
- Card 5 of 6Laws
An outside evaluator flags a foreseeable, imminent risk that a model release could kill more than 50 people. Product wants to ship on schedule.
You're the one who signs off. What's the call?
Show the answer
Answer: Block it.
Anthropic draft billUnder Anthropic's draft bill, per Politico via TechCrunch, evaluators would have 72 hours to alert Commerce, which would ask a D.C. judge to halt the release. It is a draft, not law, but no team should ship past that warning.
From the Drop: Anthropic's draft bill would let a judge halt a model release if 50+ lives are at risk
- Card 6 of 6Agents
You turn on Databricks Agent Memory, in Beta, for an internal analytics agent, with memories scoped per user and a pipeline told to skip one-off debugging notes.
You're the one who signs off. What's the call?
Show the answer
Answer: Ship it.
Data controlDatabricks says memories are scoped to each user, developers can read, export and delete them, and pipelines can be told what to keep. Its internal benchmarks showed triple the pass rate with 31% fewer tokens.
From the Drop: Databricks puts agent memory in Beta, built on its MemCurrent research
Games are written by AI from this Drop's sourced items. Scenarios are illustrative.