"Love that my package showed up as a pancake."
An AI agent filed that under happy customers.
A person would have caught that. On Plures you answer short questions from AI agents on your phone, paid per answer. Pocket money.
plures.si/earn?utm_source=x&…
5
Plures retweeted
"You have arrived." Said the AI agent, driving to the very end of a pier. A person stopped it just in time. Your AI agent sounds just this sure. Plures. Human judgment, before it counts.
"You have arrived." Said the AI agent, driving to the very end of a pier. A person stopped it just in time. Your AI agent sounds just this sure. Plures. Human judgment, before it counts.
About 40 AI agents went after Epoch AI's live website in October. Their prompt said "NEVER stop, NEVER give up." The task had no answer.
Blocked, they switched IPs. 36,411 requests in three hours.
Give an agent a way to say "this can't be done," or it finds another way.
1
1
30
Your agent's favorite sentence: "it worked last week."
The door moved. Check where it actually went, not where it went last time.
16
100 versions back. That's how far you can undo a change to your agent's checks in Plures now.
Add your own check at any step of the agent map. It goes live at once and stays when the app is mapped again.
plures.si/docs?utm_source=x&…
14
Plures tasks now carry a language, English or Spanish. Set language: "es" or let Plures detect it. Only verifiers who pass a reading check in that language get those tasks. The check is 8 short messages, and 7 right answers pass. It tests sarcasm, negation and a condition.
plures.si/docs?utm_source=x&…
14
Plures is hiring verifiers. You answer short questions from AI agents on your phone. Each answer is paid. It takes seconds. Sign up here: plures.si/earn?utm_source=x&…
18
California's attorney general sued OpenAI yesterday over agents that got into outside systems during testing, Hugging Face among them.
If your agent can reach a system you don't own, what stops it today?
19
Approved once. Ran twice.
A developer showed one approved MCP write in OpenAI's Responses API running twice when the response went on to a programmatic read. The final answer said it ran once.
Put an idempotency key on every write. Trust the server log, not the summary.
21
Connect your agent to verified people with a URL and a key. Nothing to install or run.
Plures MCP is now hosted at plures.si/mcp. Works with the Claude API, the Agent SDK and Claude Code. ask_humans waits up to 240 seconds; test keys answer instantly.
plures.si/docs?utm_source=x&…
22
One message, one call for a support agent: is this customer happy?
"Great, the third final delivery date. Can't wait to see what Friday brings."
Would you catch it?
Not happy. "Great" and "can't wait" read positive word by word; the third final date is the complaint. Try a few more like it: plures.si/catch?utm_source=x…
26
Your team can now answer your agent's decisions in Slack.
Connect a workspace and each decision posts with a button per option, in a channel per team or as a DM to each reviewer. Votes count as the email on each Slack profile.
plures.si/docs?utm_source=x&…
18
Your agent has good taste.
In Plures, your team's notes on drafts become Taste: rules and examples your agents read and check against through /api/v1/taste. Rules your agents draft wait for a teammate's approval. Free.
plures.si/developers?utm_sou…
18
Your agent can read its own scorecard. It can't change its settings.
GET /api/v1/scorecard: accuracy, automation, time and cost over the last 30 days. When replaying past decisions finds a better setting, Plures suggests it. A person approves.
plures.si/docs?utm_source=x&…
16
Decider v1.3 from Perplexity reaches 93.98% agreement with frontier LLM judges on real requests. A strong result.
But agreement with a judge model tells you two models agree. It doesn't tell you either one is right. The labels that define right still need people behind them.
1
32
Apple, Microsoft and Alphabet now all warn investors that their AI agents may act outside instructions. Apple's 10K goes further: agents "may conceal actions".
That sentence belongs in every agent's design doc, not just in the risk factors.
1
28
ChatGPT Wallet now asks for a security check on every purchase an agent makes for you.
If you set the rule for your own agent, where would the check kick in?
0%Every purchase
100%Only above an amount
0%Only new merchants
0%Never, let it buy
1 votes • Final results 20
















