AlkhemyAgent Custody

The Agent Diagnostic · free

Can you reconstruct what shaped an agent and what it actually changed?

Seven questions, one for each link from what the agent was asked to do to whether anyone would know the change mattered. About two minutes. Nothing is stored.

Or get the failure table

The chain check

Seven links, then where yours breaks first

Think of the last change an AI agent made that mattered. Answer for what you could show today, not what you could rebuild with a week of digging. No email, nothing stored.

0 of 8
Link 1 of 7. Could you show what the agent was asked to do, and what done would have looked like?

Three logged incidents from our own systems. In each, something real was happening that nothing on screen showed, until someone went looking.

16,781 to 1

Failures to successes over 70 days. A working fallback answered every query, so nothing ever looked broken.

18 hours

A dispatch pipeline dark behind a dashboard that stayed green the entire time.

177,777

Password attempts against three servers in one week, behind a default nobody had checked.

The diagnostic

Seven symptoms, the layer that owns each, the instinct to resist

Custody, who is accountable for the output between the agent claiming success and it reaching production, is the layer that decides whether you can trust the result.

Use it the day something fails: find the symptom, read the layer, and notice the fix you were about to reach for.

The diagnostic unlocks on this page the moment you submit. No spam, no sequence, one email if the cohort opens a new run.

Where you are

Seven questions about last week, then your level

Answer for what you actually did, not what you meant to do. No email, nothing stored, about ninety seconds.

0 of 7
Question 1 of 7. Last week, what did a typical request to your agent contain?

The three signals scored here are Anthropic's, from research on ~400,000 interactive sessions from ~235,000 people between October 2025 and April 2026: how precisely the user frames their directions, what they ask Claude to verify, and whether the user tends to correct Claude or Claude tends to correct the user. Read the research.

The five rung names are ours, and this is not their classifier. Theirs read real sessions. This reads seven answers you gave about your own week, so treat it as a mirror rather than a measurement. Anthropic state their own limit: As in most of our research, we cannot measure real-world outcomes, like whether code written in a session is actually used or discarded thereafter, or whether it produces an economically valuable artifact.

One finding worth the whole study if you are not an engineer: In code-producing sessions, every one of the ten largest occupations in our dataset lands within seven points of software engineers in terms of their success. This is method, not credential.

This table is the last ten minutes of Agent Custody, a four hour live cohort that builds the layer underneath it: written law, permissions that fail closed, orchestration that cannot collide, and verification that disbelieves the summary.

Request a seat

Opens an email with the subject filled in. Ten seats per run.

Muninn · an Alkhemy product

Muninn connects what steered an AI agent to what it changed.

It reads the context behind every AI-authored pull request (instruction files, memory and MCP config) and signs a content-free Context Receipt before you merge. With the Muninn Map on, it flags where that context collides with the change. Next on the roadmap: checking what a change actually does against what the agent was asked to do.