Compare
Coding agents and software factories execute the work. Methods is the control plane above that layer: it decides the right method for each decision, from intent to outcome, runs the build with its own agents underneath, and proves the result.
The layer above the factory
The market has settled into layers. Each one up decides more of what the layers below it do.
| Layer | What it does | Unit of work |
|---|---|---|
| Coding agents | Work a session or a task in the editor, terminal or cloud | A task |
| Software factories | Run the build loop: triage, spec, implement, review, open a pull request | A ticket, ending in a pull request |
| Agent control planes | Govern a fleet of agents: who each agent is, what it may do, what it costs and a log of what it did | An agent |
| Methods | Decides how each product change should be done, from the first idea to the observed result, then builds it with its own agents and proves it worked | An outcome |
Layer names as the market uses them in 2026. Vendors use the terms loosely, and a product can span layers.
The spectrum of approaches
From the least structured to the most. Each has a place; they answer different needs.
| Approach | Who drives | Unit of work | Truth lives in |
|---|---|---|---|
| Vibe coding | A person prompts and accepts output largely unread | A prompt | The chat |
| Agentic coding | An agent in the editor or terminal; a person reviews each diff | A task | The code and the conversation |
| Spec-driven development | A person approves spec, design and tasks; the agent executes | A spec | Requirements, design and task files |
| Software factories and background agents | Agents take tickets from the team's tools and open pull requests | A ticket or pull request | The tracker and the pull request |
| Methods | Methods chooses the method for each decision; agents, people and tools carry it out; a person rules on the decisions Methods will not make for you | An outcome, shaped into journeys and features with contracts | Methods, contracts, evidence and production signals |
A tool can span more than one approach, and products change.
Methods and software factories
Methods includes a factory of its own and decides how it runs, so the differences from software factories are worth stating plainly. The left column describes the approach in general; products vary.
| Software factories | Methods | |
|---|---|---|
| Unit of work | A prompt or task, ending in a diff or pull request | A feature against a contract (observable criteria, the files it may touch, verifications) inside a dependency graph |
| Who decides done | A person approves diffs by default | Automated checks decide pass or fail, tied to the latest version of the code. No output is not success |
| Adversarial testing | Review or security agents | The breaking agent: a separate agent that did not build the feature, whose only job is to break it |
| Merge | A pull request into your CI and review queue | Verified merge checks before it merges; a verification goes out of date when the code changes |
| Where attention goes | Approving plans and diffs | Answering decisions Methods will not make for you, such as changing the architecture's rules, in the console |
| Measurement | Cycle time, autonomy ratio, cost per merged pull request | Cost per completed feature, First-pass rate, Bugs found by the breaking agent, Decisions a person made |
| Where state lives | The vendor's platform | Your repository, plus a background program on your machine |
| Source | Typically closed | Open-source core, useful without an account |
| Models | Often several, routed | Provider-agnostic: Claude, OpenAI, local models, deterministic workers and humans |
| Method | Whatever the team already does | Ships a method that covers the whole path from idea to learning from production, with standards turned into automated checks |
When another approach is the better fit
- For a throwaway prototype, vibe coding is faster and the cost of a mistake is low.
- For a quick edit inside a codebase you know, an agent in your editor is enough.
- If your CI and human review already carry the quality bar, a background agent that opens pull requests may fit as it is.
- Methods is for teams that want a feature proven against criteria fixed beforehand, with the evidence attached, and who accept a stricter process to get it.
Build with us as a design partner.
Methods is pre-alpha. A small number of teams will run it on real work while it is built, and shape what it becomes.