A doctor friend asked me about billing agents their mixed experience not being able to dispute the agent due to abstractions. Wanted your thoughts. I feel agents should prove they are trustworthy and transparent to the human in the loop as in the end user, not the developer.

I built a tool where every AI agent action gets an Ed25519 signed receipt, hash-chained, and exported as a folder anyone can verify with openssl. No vendor software needed to verify.

The demo runs 20 simulated insurance claims sessions — standard agents and rogue agents. Every action is signed. Rogue actions are blocked and the blocks are signed too.

$ pip install agentmint && python -m agentmint.demo.healthcare

Cross-patient access: 10 attempts → 10 blocked Auto-deny (no review): 10 attempts → 10 blocked Prompt injection: 10 attempts → 10 caught before LLM

To run it: clone then cd agentmint-python/ Verify: cd healthcare_evidence && bash VERIFY.sh GitHub: https://github.com/aniketh-maddipati/agentmint-python/tree/m...

Also Would love to hear: Should verification require zero dependency on the vendor, or is that overkill?

Tear it apart, tell me what I can do to make it better. If you have thoughts on ways to make trust a currency for agents, comment below

I'm not sure I understand the point of this; is the intended user of this an insurance company whose agents approve or deny claims?

If so, is there any actual incentive for them to allow doctors & patients to follow up with this sort of paper trail?

Both, the insurer installs this to produced portable evidence that end users can consume. Would mainly be the startups building AI agents that interact with insurers on behalf of hospitals. If a hospital needs to process 100,000 claims made by agents and verify them, its their records pitted with vendor logs. This lets the vendor hand over a folder the hospital verifies independently with openssl. No vendor trust required. The incentive: customers (clinics) can process deterministic evidence independent of the vendor or tool (my code) and the vendors agents build up a track record of trust by themselves without engineers spending their time triaging.

1.2 seconds is fast enough that providers need a different mental model. You cannot read the denial reason before the next one arrives. The more interesting question is whether the counterparty should be an agent too. At empla.io founders forward denied invoices, disputed charges, or auto-rejected vendor requests and the agent negotiates back inside the thread. Most denial infrastructure was built assuming a human on the other end. That assumption is about to break.

Agent-vs-agent makes the trust problem sharper, not necessarily easier. Each side needs verifiable evidence of what the other committed to. They cant do with just vendor logs. The "one side controls the record" assumption breaks right after the human-on-the-other-end one. Plus, its like waiting for the police after an incident with out a first-aid kit.