Foundry for real projects

Your agents should remember your project.

Foundry watches the work you already do, captures the corrections and decisions that matter, and turns them into durable context for the next run.

Cyberpunk foundry with a luminous grand program under construction.Greek workshop with a monumental bronze titan-scale project under construction.Arkham drafting hall with an impossible architectural system being stabilized.
Visual system Your project, staged as serious craft The language stays practical. The imagery treats ordinary project work as something worth building carefully. Cyberpunk / Greek / Arkham
The promise

Every run should leave the project a little easier to align.

Foundry is not trying to make the model mystical. It gives your project a memory system, a routing layer, and a way to verify whether the corpus is actually helping.

Capture what matters

Corrections, decisions, conventions, review feedback, and repeated explanations stop disappearing at the end of a session.

Route it somewhere useful

A signal can become memory, documentation, a fixture, a guardrail, an ADR, or a CI rule. Foundry helps decide which.

Make the next run better

The point is not more context. The point is the right context, carried forward and verified against real work.

The team around the work

Each role contributes a different kind of project memory.

The visual story is heightened; the jobs are concrete. Map the relevant context. Capture the lesson. Watch parallel work. Protect project boundaries. Do the task.

around the project CARTOGRAPHER
CARTOGRAPHER — Maps contextCARTOGRAPHER — Maps contextCARTOGRAPHER — Maps contextLIBRARIAN — Captures learningLIBRARIAN — Captures learningLIBRARIAN — Captures learningHERALD — Watches the wireHERALD — Watches the wireHERALD — Watches the wireWARDEN — Guards boundariesWARDEN — Guards boundariesWARDEN — Guards boundariesARTIFICER — Does the workARTIFICER — Does the workARTIFICER — Does the work

CARTOGRAPHER

Maps context

Finds the project context a task actually needs, so the working agent does not have to carry the whole corpus.

The loop

The next attempt should start closer to what you mean.

Not more autonomy for its own sake. Better starting conditions, better feedback, and fewer lessons paid for twice.

Work normally

Use Claude Code, Cursor, or your existing agent workflow. Foundry wraps the session instead of replacing it.

A signal appears

A correction, redirect, revert, or review comment says something important about how your project should work.

Foundry classifies it

The Librarian decides whether it is taste, convention, security, architecture, fixture material, or something else.

The corpus changes

Approved signals become durable project context your agents can use next time.

Oracle checks it

The same kind of task runs again against fixtures, with scores and regressions tracked.

The lesson compounds

If it helped, the improvement stays. If it hurt, you know before it becomes invisible drag.

Foundry + Oracle

Build in the workshop. Hone it with measurement.

Start by making project signals durable. Add Oracle when you want to know whether those corpus changes actually improve future agent work.