Trust is the bottleneck.
Agents can finally do real work. Real work means taking consequential actions: refunding customers, rolling back deploys, updating the CRM, merging code. But agents make mistakes, just like we do.
That gives us two options:
- Let agents act, and accept that some actions will go wrong.
- Don’t let them act at all.
Most choose the second, because permission today is a switch. An agent can issue refunds or it can’t, and a $10 refund looks the same as a $10,000 one. So we don’t let agents issue refunds at all. The bottleneck is no longer what agents can do. It’s what we let them do.
There’s a third option: let agents take consequential actions, with a human in the loop. The $10 refund goes through on its own. The $10,000 one waits for the right person to approve it.
Agents are also increasingly autonomous and distributed. Support agents in your helpdesk, coding agents on developer laptops, accounting agents run by a vendor. Each one comes with its own rules, or none at all.
The rules belong in one place your organization owns, not in every agent. Support sets the rules for how refunds get approved, regardless of the agent issuing the refund. Sales owns the CRM. Engineering owns production.
That’s what we’re building at withHuman: an open standard for approvals that any agent harness can use, and an open-source platform where your team decides what runs automatically and which humans approve the rest.
Read more at our launch post.
The withHumans
Chris spent three years at Palantir building the compute platform behind Foundry. After Palantir he co-founded Metoro, an AI SRE Agent that fixes production issues on its own. He was giving agents safe access to prod before it was cool.
Follow Chris
Ece spent four years at Amazon keeping Prime Video reliable for a hundred million customers, then co-founded Metoro. She claims she gave Metoro's agent access to prod first, but she's not writing this so we'll never know.
Follow Ece