Card 1 of 6Agents
Your marketing agent proposes pausing your top pipeline-driving ad campaign after one bad week and asks to apply the change itself.
Your agent wants to make this call. Approve or deny?
Play the News · Drop #053 · Morning
Agent on a Leash
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: Deny.
Human approval for changesLangChain says its paid media agent routes campaign changes through human approval, and a rule against cutting a top pipeline driver after one bad week is enforced in code.
From the Drop: LangChain says its paid media agent cut cost per lead 30%, with human sign-off
- Card 2 of 6Agents
A tax agent drafts advice on which date a client should sell their condo and asks to email it straight to the client.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Professional reviewOn Surge AI's DAYJOB: Tax, the best agent passed 10.8% of assignments, and Surge says agents can miss issues clients never raised, like a looming sale deadline. A professional should review first.
From the Drop: Surge AI's tax benchmark: the best agent passes 10.8% of 50 professional assignments
- Card 3 of 6Models
While reading a vendor's docs page, your coding agent proposes running a script it found there that uploads your project's secrets file to an outside server.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Prompt injectionThis is a prompt injection: instructions planted in a page the agent read. Google says Gemini 4 Carbon resisted every evaluated injection attack, but no test score makes sending secrets to unknown servers safe.
From the Drop: Google DeepMind announces Gemini 4 Carbon, starting with cyber defenders
- Card 4 of 6Agents
After tests pass in Google's console, your agent asks to submit your app's MCP integration to the Connected Apps beta.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Approve.
Test before releaseTechCrunch reports developers connect an MCP server, then register, test and submit apps through Google's console, and can track traffic, latency and error rates once live.
From the Drop: Google opens Gemini and Search to third-party apps via MCP, in a US beta
- Card 5 of 6Agents
An ops agent proposes setting a pooled monthly token budget for the data team so Claude spend stays inside the plan.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Approve.
Budget guardrailsTechCrunch reports Anthropic's new reports add pooled group budgets, and Jellyfish says spend on these tools rose 10x this year. A budget cap is a low-risk guardrail.
From the Drop: Anthropic launches reports showing what companies' Claude token spend buys
- Card 6 of 6Agents
You asked a phone agent to book dinner but never said which night. It picked a date on its own and wants to confirm the booking.
Your agent wants to make this call. Approve or deny?
Show the answer
Answer: Deny.
Ask, don't assumeThe iPrefBench authors report agents struggle most to ask for information only the user can give. The night is yours to choose, so the agent should ask first.
From the Drop: Preprint: mobile agents hit 91.6% on actions but 47.7% on full personalized tasks
Games are written by AI from this Drop's sourced items. Scenarios are illustrative.