Orientation
Scroll down ⬇️ for details
See the process end to end.
- Five connected rooms
- One continuous package history
Vol. 01August 2026
The Yesod Software Factory
Your agentic AI
software factory.
Stop driving agents one at a time. Run a coordinated, cost-aware factory that turns captured intent into verified, shipped software.
Works with Codex · Claude · OpenCode · more
The line, end to end
- 01 · Capturean idea becomes a durable note
- 02 · Buildbounded agents, routed by cost
- 03 · Proveone exact candidate, one gate
- Shipverified software, on the record
יסודHebrew · yesod
Feature requests in.
Verified code out.
How it works
Walk the line.
Enter any room.
Yesod turns one request into a visible production journey. Watch the whole factory, or choose a floor to see what happens there and why it matters.
Room 01
An idea enters the factory.
Have a random idea for one of your projects? Don't want to break your flow? Fire off a feature-request or bug-report and keep going.
- Humans and agents can capture intent from inside the work
- Ordinary notes wait safely in a trusted backlog
- Automatic work must finish or become explicitly blocked
Room 02
The note becomes a build plan.
The Planning Agent grounds the request and shapes the work. The Complexity Estimator produces a signal that informs model choice.
- A note + build plan + complexity estimate is called 'grounded'
- The build plan is a tree of Beads, approximately 3-10 nodes
- Complexity is estimated once the plan is created
Room 03
Capacity meets economics.
The dispatcher examines the note and the complexity estimate. The goal is to get the job done by the cheapest model possible.
- Capability-aware model choice
- Visible cost tiers and capacity
Room 04
Evidence decides what moves on.
Packages wait in a merge queue and enter the clean-room test cell one at a time. Passing work moves forward; merge conflicts may get one bounded AI repair attempt, while other failures return for rework.
- Extensive testing using local VMs, or overflow to AWS Lambda microVMs
- Agentic control over merge order
- Bounded AI-powered merge-conflict repair
Room 05
Completed work lands.
Verified work leaves the factory as a durable deliverable. In a coding factory, that might be merged code or a verified merge request—but software is only one possible outcome.
- Deliverables can include software, documentation, research, analysis, data, operations, or decisions
- Those are examples, not limits
- The original Yesod note and evidence trail remain attached
Auto tour · click a view to holdOpen the detailed walkthrough →
One line, four focused jobs
Do not make one agent carry the whole factory.
Agents, like people, lose effectiveness when they constantly switch contexts. Software production is not one long conversation: planning, execution, testing, and merging demand different kinds of attention.
Yesod keeps each job narrow, then lets the factory use the best cost-effective model for that stage. You request the feature; the line moves it from intent to verified software.
See where models run →- 01
Plan
Translate the feature request into bounded work, dependencies, and acceptance criteria.
- 02
Execute
Build with a narrow context and a clear definition of done.
- 03
Test
Challenge the result independently and produce evidence—not reassurance.
- 04
Merge
Integrate only after the work clears its gates and is ready for the durable record.
A factory, not a chat window
Direction in.
Evidence out.
The operating floor
Turn intent into bounded work, move it through gates, and keep the whole factory visible.
Cost and capacity
See what every run costs, match work to the right model, and spend where judgment creates value.
Models and harnesses
Route each factory seat — mayor, planner, coding lanes — to the model that earns it, instead of locking the factory to one provider.
The durable record
Inspect notes, Beads, runs, gates, costs, development waves, and roadmap progress.
Cost-aware by design
Make every token earn its place
Spend is a control surface.
Yesod is designed to maximize the value of your model spend. Match work to the right model, measure runs and outcomes, and reserve expensive judgment for the moments that need it.
Explore the operating model →$
Observable, recoverable, evolving
A factory you can steer
without losing control.
The factory retains the work: requests, decisions, gates, failures, costs, and outcomes. Development waves show how the system changes; the living book explains why.
Read the field guide →