For software developers
Hand off bounded work without losing architecture, tests, or review discipline. Guildhall nudges agents toward existing code, shared components, focused verification, and visible evidence.
Skip to content
AI agents for real software work.
Give AI agents a job site,not just a chat box.
Plan the task. Run the agents. Review the evidence. Recover when work gets stuck.
- Feature slices, cleanup passes, docs alignment, and release prep.
- Visible tasks, blockers, transcripts, reviews, and gates.
- Enough structure to keep AI work from drifting into chat fog.
Current docs: 0.8.0.
- Shared agent state
- Blueprints before changes
- Reviews before done
Why use it?
Because AI work gets better when the project has a shape.
For product-minded builders
Explain the product clearly, then let Guildhall turn that intent into smaller task blueprints. You can answer scope and taste questions without pretending to be the compiler.
For messy real projects
Keep work from scattering across chats, half-runs, and “wait, what did it change?” moments. Guildhall keeps the trail attached to the task.
Guildhall borrows its name from a shared room for skilled work: different trades, common standards, visible progress. That is the product promise, minus the dust.
The core idea
An agent harness wraps the model with workflow.
A chat assistant answers a conversation. An agent harness decides what work exists, what context each agent gets, which tools are allowed, when review happens, and when the run should pause for you.
Plan
Turn broad intent into a blueprint: goal, scope, non-goals, acceptance criteria, and checks.
Build
Give workers focused context so they can make changes that fit the project instead of inventing a parallel universe.
Inspect
Use reviewers and gates to attach findings, command output, and release evidence before calling work done.
First steps
Read just enough, then try one small task.
The best first run is intentionally modest: one project, one task, one visible path from idea to review.