CodeMachine

You will get value from this the third time you run it, not the first. The pitch is that the sequence you normally hold in your head becomes something written down and repeatable, so the reproduce-analyse-plan-implement-test rhythm survives being tired at four in the afternoon. That repeatability is the deciding trait, because consistency is the thing humans lose first and machines never had.

Pick it if you do the same shaped work across many repositories. Pick Claude Code alone when a single agent with its own subagents already gets you there without a second layer to maintain.

The coupling is the problem. This spawns third-party agent binaries and steers them through scripting modes those vendors maintain for their own reasons, none of which include compatibility with an orchestrator they do not know about. When one changes an output format, the failure surfaces inside a long run rather than at startup, and a run designed to last for days will have consumed a great deal before anyone notices.

What it does right: state is persisted across a run, so a failure resumes from where it stopped rather than starting the whole sequence again.

1. The context discipline is the strongest idea here: prompts live in one place and each step receives a scoped view rather than the accumulated transcript, which is the correct answer to the failure that ruins long agent runs. 2. Different agents are assigned to different stages, so a step's model can match its difficulty. 3. Verification appears as an explicit stage rather than an implicit hope.

No benchmark accompanies any of it and the capability description is a single marketing page. The observation: the architecture is more thoughtful than the documentation supporting it.

Traction is decent for a year-old project and the position is structurally weak. Orchestration above other people's agents is the first thing those agents absorb, because every vendor beneath this layer is actively shipping their own delegation, scheduling and parallelism. There is no price, no hosted product and no data asset, so nothing accumulates while the ground shifts.

Likely path: absorbed by feature parity rather than acquired, since there is little to buy beyond the workflow format. Position: use it, do not depend on it, and expect the capability to arrive free inside a tool you already run.

One sentence on the demo: it worked overnight and produced a branch. The commercial exposure is not the tool, which costs nothing, it is that this thing exists to run several paid agents at once for extended periods, so our inference spend becomes a function of how ambitious an engineer felt on a Friday. There is no console, no identity integration and no central record of what ran or why.

Onboarding is a day for anyone who already uses the underlying agents. Not yet, absent a spend control we administer.

The licence and the install are both what I want: permissive, one command, no account, nothing phoning home. The workflows are definitions on disk, so they live in a repository, diff properly and travel between machines, which is the difference between a tool I configure and a service I subscribe to.

Two gaps annoy me. There is no tool-protocol client here, so my servers only reach the agents underneath rather than the orchestrator, and no local model runtime is documented, so this layer assumes hosted engines throughout. Both are patchable, and the licence says so.