Make Your AI Agent Verify Its Own Work — No More Hallucinated "Done"

July 2, 2026 howto hermes v0.18.0 verification

When I ask my agent to fix a bug or write a module, I don't want it to say "done" — I want it to prove it. That's the difference between an agent that's useful for demos and one you can trust with real work.

Hermes Agent v0.18.0 ("The Judgment Release", July 1, 2026) introduces a verification system that changes how agents finish tasks. Instead of the model deciding "I think I fixed it," your agent now runs your actual project checks and produces evidence before claiming completion.

"The tests pass. Here's proof." — Instead of "I think it works."

The Trust Problem with AI Agents

Every business I talk to who's tried deploying an AI agent runs into the same wall: how do I know it actually did what it said? LLMs are optimised to sound confident, not to be accurate. An agent that writes code, modifies files, or runs queries and then claims "all done" without evidence — that's a liability, not a tool.

Before v0.18.0, an agent's standard for "done" was its own internal certainty. If you asked "is the API endpoint working?" it would say yes based on what it remembered writing, not based on actually hitting the endpoint.

What Changed: Completion Contracts

Every /goal directive can now include a completion contract — a statement of what "done" looks like, measured by real evidence. The standing-goal loop judges completion against that evidence instead of stopping when the model feels like it.

Here's a practical example. Before:

/goal Add login validation to the API

# Agent writes code, says "done"
# No proof. Maybe it works. Maybe it doesn't.

After — with a completion contract:

/goal Add login validation to the API
done when: the test suite passes, the new endpoint
returns 200 for valid credentials and 401 for invalid ones,
and a curl smoke test confirms both cases

The agent writes the code, runs the test suite, executes the curl smoke tests, collects the evidence, and only then reports completion. If any check fails, the agent tries again or reports what broke.

How Verification Evidence Works

Behind the scenes, the agent now records verification evidence for every significant action:

Test Suite Runs

The agent executes your project's actual tests (pytest, vitest, whatever you use) and captures the output.

SMOKE CHECKs

Curl endpoints, check HTTP status codes, validate response bodies. Real network calls, not speculation.

File Verification

Stat files, check contents, run linters. The agent confirms files exist with the right content.

Custom Hooks

pre_verify hooks let you wire in any custom check. Run your deployment pipeline or integration tests.

Why This Matters for Your Business

Here's the blunt truth: you shouldn't deploy an agent that can't verify its own work. Not for customer-facing code, not for internal tooling, not for anything that costs you money if it's wrong.

The verification system in v0.18.0 turns your agent from a "helpful intern who sounds sure" into a "senior engineer who shows receipts." You still review — but you review evidence, not promises.

Three business scenarios where this matters immediately:

Revenue-critical code changes — An agent modifying a checkout flow must prove the payment gateway responds correctly before claiming done
Data pipeline automation — An agent transforming CRM exports must verify every record processed matches the expected count
Deployment readiness — An agent preparing a release must run the full CI suite and show you the green checkmarks

Getting Started

If you're on Hermes Agent v0.18.0+, the verification system works out of the box with sensible defaults. No config required to start. The one-time migration tunes the defaults for your project:

hermes upgrade # to v0.18.0 or later
# Verification is active by default.
# Start a goal with a "done when" clause and watch the evidence roll in.

The pre_verify hook is available if you want custom pipelines. Full docs are in the release notes.

Pro tip: Start with small scope — "done when the endpoint returns 200" — and expand to full test suites as you build trust. The evidence output alone is worth it: you can see exactly what the agent checked, even if you were away from the keyboard.

What Else v0.18.0 Ships

This release is massive. Besides verification, it also ships:

MoA as a first-class model — Pick an ensemble of frontier models like any other model. See every model's reasoning, then watch the answer stream in.
/learn — Turn any directory, URL, or workflow into a reusable skill with one command.
/journey — A playable timeline of everything your agent has learned. Edit or delete memories and skills right from the view.
Background fan-out — Run multiple subagents in parallel while you keep working. Results come back consolidated.

But honestly? The verification system is the one I'd start with. Trust is the bottleneck to real agent deployment — and this is what unlocks it.

Try It Yourself — First Month Free

Deploy a Hermes Agent with the new verification system. Use coupon code blog950 for your first month free — no risk, full access to v0.18.0.

Get Your Agent →

derez.ai — Deploy your AI agent in 5 minutes. · Setup Guide · v0.18.0 Release Notes