How Railspike compares
Railspike (railspike.dev) isn't really in the same category as most of the tools above. Those decide which agents run, where, and in what order. Railspike decides what one agent is allowed to do and when a person signs off. The agent writes each stage as a Markdown file in docs/plans/ (Assess, Define, Plan, then Build one stage per commit), you edit and approve it in your own editor, and the approved file becomes the prompt.
The closest match is Mastra's factory, which also has three human gates (accept the task, approve the plan, merge). Railspike's three gates all come before any code is written. MetaGPT and ChatDev also pass documents between stages, but there the agents review each other. Railspike puts a person back at that step. StrongDM shares the idea that the written spec is the real input and then goes the other way: StrongDM keeps humans off the code and trusts hidden scenarios, while Railspike keeps a human on every plan.
It has little in common with Factory.ai or Gas Town, which are built on running many agents at once. It also doesn't compete with Composio, Emdash, Claude Squad or Bernstein, which handle worktrees, queues and dashboards. A Railspike plan could just be what each of their worktree agents follows.
What it does that most of them don't: the state is plain files in the repo, so you can lose context, start a new session, or switch from Claude Code to Codex partway through without losing anything. It's close to dependency-free, installed by pasting a prompt. And because the reviewed plan is the prompt, you approve exactly what the agent will act on.
Where it's thinner: verification. Stages include checks, but nothing independent grades the result, and verification is the part the whole field is stuck on. The Define step ("what done means") seems like the natural place for StrongDM-style acceptance scenarios that the Build agent never sees. It also doesn't try to parallelize work or run unattended, except in one-shot mode. That's deliberate, since stopping drift is the point, and it's the clearest line between Railspike and anything that calls itself a factory.
On Shapiro's ladder it's a level 3 practice, delegation with a person approving every plan. Taskplane's phrase "light factory" fits it better than "factory" does.
https://railspike.dev
#ai #agents #softwarefactory #railspike
#
Railspike (railspike.dev) isn't really in the same category as most of the tools above. Those decide which agents run, where, and in what order. Railspike decides what one agent is allowed to do and when a person signs off. The agent writes each stage as a Markdown file in docs/plans/ (Assess, Define, Plan, then Build one stage per commit), you edit and approve it in your own editor, and the approved file becomes the prompt.
The closest match is Mastra's factory, which also has three human gates (accept the task, approve the plan, merge). Railspike's three gates all come before any code is written. MetaGPT and ChatDev also pass documents between stages, but there the agents review each other. Railspike puts a person back at that step. StrongDM shares the idea that the written spec is the real input and then goes the other way: StrongDM keeps humans off the code and trusts hidden scenarios, while Railspike keeps a human on every plan.
It has little in common with Factory.ai or Gas Town, which are built on running many agents at once. It also doesn't compete with Composio, Emdash, Claude Squad or Bernstein, which handle worktrees, queues and dashboards. A Railspike plan could just be what each of their worktree agents follows.
What it does that most of them don't: the state is plain files in the repo, so you can lose context, start a new session, or switch from Claude Code to Codex partway through without losing anything. It's close to dependency-free, installed by pasting a prompt. And because the reviewed plan is the prompt, you approve exactly what the agent will act on.
Where it's thinner: verification. Stages include checks, but nothing independent grades the result, and verification is the part the whole field is stuck on. The Define step ("what done means") seems like the natural place for StrongDM-style acceptance scenarios that the Build agent never sees. It also doesn't try to parallelize work or run unattended, except in one-shot mode. That's deliberate, since stopping drift is the point, and it's the clearest line between Railspike and anything that calls itself a factory.
On Shapiro's ladder it's a level 3 practice, delegation with a person approving every plan. Taskplane's phrase "light factory" fits it better than "factory" does.
https://railspike.dev
#ai #agents #softwarefactory #railspike