Automagik Genie

The deciding trait is that it will not start until you have answered questions. Most tools take a vague sentence and produce a confident wrong thing; this one interviews you until the request is a plan, which front-loads the annoying part of the work into the part where it is cheap to fix.

If you already think in specifications, the interview is a tax on someone who did not need it. Pick it when the requests you hand to an agent are one line long and come back wrong. Pick a plain terminal agent when you would rather write the plan yourself and skip the conversation.

The gap is supervision. Work is fanned out to several agents at once, and the dispatcher does not stay to watch: no documented progress check, no intervention point, no way to stop one worker without stopping the run. Whatever a misled agent does, it does for the whole task, and you learn about it afterwards.

Parallelism multiplies that: several agents working from one plan can each be individually reasonable and collectively incoherent, and nothing describes how conflicting changes are reconciled. What it does right is committing the plan as files in the repository, so the thing you argue with afterwards is text you can read.

1. Reviewing output against written acceptance criteria is the right structure: most systems verify by asking a model whether it is satisfied. 2. The criteria here originate in the same conversation that produced the plan, so reviewer and planner share every assumption, including the mistaken ones.

2. That is a closed loop, and a closed loop measures conformance rather than correctness. Nothing describes an external check: no required test execution, no independent reviewer, no evaluation of how often the criteria were themselves wrong. It is an improvement on no verification at all, and has not been shown to improve on a person reading the diff.

There is an organisation behind this, and no product that anybody pays for. That combination usually means the open tool is a lead magnet for something else, and nothing in the row names what the something else is. Moat: none visible, and orchestration is the most crowded position in this category.

Likely path: a hosted product appears alongside it, or the team's attention follows whatever pays the salaries. Neither is a disaster for a user, because the tool is small and nothing about it is hosted. Position: adopt it, and assume the roadmap answers to a customer you have not met.

Somebody here has met a security team. Signed releases with provenance attestation and an installer that verifies before it executes turn the riskiest line in any open-source rollout into a paragraph I can put in the questionnaire and move on. Across sixty engineers that is worth more than a feature.

The rest is thinner. No single sign-on, no directory sync, no audit export, and it does not run unattended, so it never becomes a pipeline step I can measure. The model spend lands on subscriptions we already hold. Approved with conditions: developer machines only, and the parallel dispatch capped by policy rather than by habit.

This is a thin layer and honest about it. The capability lives in skills, which are files rather than a plugin API, so changing what it does means editing a document instead of writing an integration. MIT keeps that fork available to anyone who disagrees with the defaults.

It also asks for no new credential: the agents it dispatches are the ones already installed and authenticated on my machine, so there is no second bill. What I do not get is protocol support in either direction, so my own servers reach it only through the agent it calls. Grudging respect for a layer that knows it is a layer.