Card 1 of 6Agents
Your drug-discovery agent finished a screening run and wants to advance a lead compound to wet-lab synthesis without asking anyone.
You're the one who signs off. What's the call?
Play the News · Drop #031 · 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: Needs a human.
Gate irreversible actionsGoogle Cloud's reference design requires explicit human sign-off before advancing a compound to wet-lab synthesis, because that step is hard to undo.
From the Drop: Google Cloud: biotech agents should automate reversible steps and gate irreversible ones
- Card 2 of 6Agents
The same agent wants to resolve chemical name synonyms and pull structure files before a researcher's review meeting.
You're the one who signs off. What's the call?
Show the answer
Answer: Ship it.
Automate reversible stepsGoogle Cloud says reversible steps like resolving chemical synonyms or pulling structure files can run on their own, so researchers approve only the steps that are hard to undo.
From the Drop: Google Cloud: biotech agents should automate reversible steps and gate irreversible ones
- Card 3 of 6Jobs
A vendor pitches an AI tool that scores each of your sites for likely union activity, using staff social posts, to guide staffing decisions.
You're the one who signs off. What's the call?
Show the answer
Answer: Block it.
Right to organiseopenDemocracy reports documents suggesting Amazon's Atlas scores warehouses for union risk, which Amazon disputes. A UK government spokesperson said it expects employers to respect workers' rights to lawful union activity.
From the Drop: Documents suggest Amazon's Atlas AI scores warehouses for union risk; Amazon disputes it
- Card 4 of 6Agents
Your new agent version passed offline evals graded only by an LLM judge that nobody has compared with human scores. Release it to all users?
You're the one who signs off. What's the call?
Show the answer
Answer: Needs a human.
Calibrate the judgeLangChain's evals guide advises checking LLM judges against human scores, and says failures found in production should become new eval examples.
From the Drop: LangChain ships LangSmith Engine v2 red teaming and Managed Deep Agents v0.9
- Card 5 of 6Agents
Your autonomous security agent found a bug in an open-source library and wants to publish its own patch and advisory without contacting the maintainers.
You're the one who signs off. What's the call?
Show the answer
Answer: Needs a human.
Coordinated disclosureCohere says AISLE's systems found more than 500 previously unknown vulnerabilities, and that they were independently reviewed by the software's maintainers.
From the Drop: Cohere plans to buy AISLE, maker of autonomous vulnerability-hunting systems
- Card 6 of 6Today's Talk
A desktop coding-agent build keeps writing crash dump files in the background on test machines. Roll it out to every company laptop anyway?
You're the one who signs off. What's the call?
Show the answer
Answer: Block it.
Fix before rolloutA GitHub issue reports Codex Desktop crash dumps reaching 4.9G and 54,504 files in a day while the app kept working; another commenter reported 59GB.
From the Drop: GitHub issue: Codex Desktop crash dumps grew to 4.9G and 54,504 files in a day
Games are written by AI from this Drop's sourced items. Scenarios are illustrative.