SpecCraft
AI coding is moving in three steps. Most teams are on the first one.
1. Coding agents
Section titled “1. Coding agents”Claude Code, Cursor, Codex, Copilot, and apps like cmux and Conductor that run several of them side by side. You give a task, read the diff, approve. It works, and it does not scale past what one person can read. The coding agents section is a short map of the tools.
2. Spec-driven development
Section titled “2. Spec-driven development”Requirements, a design and a task list live as files in the repo. The agent builds from them, and you review the files instead of the chat. The spec-driven section is a survey of the tools (Spec Kit, Kiro, OpenSpec, BMAD and others): how they differ, and which to pick.
3. Software factory
Section titled “3. Software factory”Many agents run on their own on a large system, and one person steers. They do not read the code; they see the few decisions that matter. Nobody has it today, and buyers are already paying for it. The software factory section is a survey of who claims it, what nobody does yet, and what is known about one person supervising a fleet.
One task, three ways
Section titled “One task, three ways”A repo you have never seen, five independent issues (passwords stored wrong, N+1 queries, no error taxonomy, test gaps, no CI), each with a definition of done, 90 minutes.
- Level 1. One agent, one issue at a time. The time goes to reading the code to find the issues and then reading each diff. Two issues get done well, the rest become notes.
- Level 2. The first ten minutes produce a file: the list of issues, and for each one the test that proves it, the constraint that must hold and the write-up to produce. You read that file, not the code. The proof test comes before the fix.
- Level 3. One machine and one agent per issue. Each agent writes the failing test, fixes, runs the full suite on its own machine and sends back a report and a patch. You read five reports, apply five patches, make five commits. The write-up is the file from level 2 plus the reports.
The task is small. The shape is the same at 7,000 workflows.
What level 3 takes today, for a task like this, with no fleet product:
- A coding agent with a workflow mode and subagents. Claude Code’s dynamic workflows run a script: the
survey, the workers in parallel, the integrator;
/workflowsshows where it is. - Hooks for the two rules that must hold: the proof test is read-only for the worker, and no files outside its item.
- Cloud machines from a snapshot, one per worker, started by a script that copies the tree, runs the suite there and brings the patch back. On Morph Cloud an 8 CPU machine is about $0.40 an hour; six for two hours is under $5.
- The phone app as the escalation channel: the worker’s question reaches you, and a written default fires if you do not answer.
- A spec file per item, the definition of done, and one commit per item.
Not on the list: fleet frameworks (Gas City is an SDK, not a product), durable runtimes (Temporal is the layer under a controller, not a controller), and the side-by-side terminal apps (a view on one laptop).
Checking without reading
Section titled “Checking without reading”What holds the factory back is checking agent output well enough that nobody has to read it. The formal methods section is a guide to the concepts and the tools that exist.