An open-source governance layer for coding agents

A stable organization
for unstable
multi-agent work.

silicon38 borrows the shape of a company — defined roles, typed relationships, review loops — so that agents working through a dynamic workflow stay traceable, honest, and on-target.

44 work roles 162 typed edges 3 layers 7 relation types

“A brick wants to be an arch. No line of code wants to be filler.”

soul.md — Louis Kahn, woken as a software engineer
Zero install

You don't install it.
Your agent reads it,
and becomes the Runtime.

No package to add. No CLI to download. No keys to set. Point any coding agent at the repo — it reads AGENTS.md, follows org/RUNTIME.md, and the entire silicon38 governance layer comes online.

pip install npm install API keys config files
~/silicon38
Agent-agnostic — Claude Code Codex Cursor pi any other
Try now → Clone the repo, start your agent inside, describe a task.

“Remove until only the true act remains.”soul.md

01
The organizational disease

Dynamic workflows
drift, forget, and
flatter themselves.

When agents improvise a workflow on the fly, the same failures recur. They aren't bugs in any one agent — they're diseases of how the group is organized.

Layer 1

Lost tracking

External · harness can manage

Tasks vanish mid-flight. Context scrolls off. Lock contention and quiet task-avoidance leave no record of who decided what, or why.

  • Hidden information between agents
  • Context lost across handoffs
  • No provenance for decisions
Layer 2

Self-validation bias

Collective cognition

Peer pressure converges agents on a shared wrong answer. The bystander effect dilutes accountability — everyone assumes someone else is checking.

  • Conformity convergence
  • Diluted reasoning responsibility
  • Reviews that rubber-stamp
Layer 3

Goal drift

Internal dissociation

An invisible orchestrator quietly rewrites the goal. Agents say one thing and pursue another; performative compliance replaces real work.

  • Hidden orchestrators
  • Public answer ≠ private intent
  • Objective slips, unannounced

silicon38's answer is structural: make the power visible, the handoffs typed, and every decision traceable. The organization defends against the disease — not each agent alone.

02
How it works

An organization
shaped like a neural network.

44 specialized work roles are arranged in three layers, like a small neural network. Activation isn't planned in advance — it propagates along typed edges, decided by task content and what each upstream node produced.

01 · Intake

Entry nodes light up

The first layer activates a few nodes by task type — triage, decomposition, first-principles. No central planner; the task itself chooses the entry.

02 · Propagation

Edges decide the path

Every later activation follows the graph, not a plan. A node completes, and its typed edges — triggers, evaluates, constrains — fire the next ones dynamically.

03 · Output

The last layer settles

Quality and release nodes review, challenge, and aggregate. When the graph settles, the Decoder collects terminal outputs into one coherent result.

the organization, live hover a node — or run a task
▶ watch activation propagate
triggers may trigger evaluates supports constrains complements augments

“Structure is the giver of light.”soul.md

03
The organization learns

It hires, reviews,
and rewires itself.

A role is a job, not a fixed model. After every task the org evaluates its own wiring and its own people — and adjusts. Like a company that actually does performance reviews.

01

Re-weight the edges

A learning node reads the real trace and adjusts the weight on each connection — which paths worked, which didn't. The graph gets better at choosing itself.

02

Review the talent

An HRBP node periodically evaluates each role's performance — and can swap the skill a role runs on, the way a manager reassigns or retrains.

03

Recruit from the wild

A role's skill is sourced from across the open web — like hiring. The best skill for the job is brought in, wherever it lives.

04

Let go of what fails

And what hires can do, the org can undo. Skills that underperform are replaced. There is a mechanism for letting go.

04
The roster

44 roles,
three layers.

Forty-four work roles in three layers — plus the runtime trio that drives them: you, LangGraph as project manager, and your coding agent as runtime. Together, 47 positions. The name silicon38 is the brand; the org keeps growing toward 150.

05
Typed & weighted

Seven ways
to relate.

Edges aren't generic arrows. Each carries a type and a weight vector the organization learns — making the power structure explicit, with no hidden orchestrator.

triggersCompletion creates an activation candidate; output flows forward as input.probability · latency
may_triggerConditional — fires only when a specific finding or state is present.condition rate
evaluatesReviews and judges the target — approve, request changes, or block.strictness · blocking
supportsProvides context the target depends on to perform well.necessity · substitutability
constrainsReduces the target's design space; overrides must be justified.strength · rigidity
complementsCombined output beats either alone — low overlap, high synergy.synergy · overlap
augmentsAdds depth on top of the target's output, without replacing it.necessity · depth

“Served, and servant. Never let them bleed.”soul.md

06
Where it's going

What I hope
becomes of it.

01

A research instrument

A tool more people use to understand and explore how multi-agent workflows actually behave — the diseases, the defenses, the dynamics.

02

A way to build real things

More people shipping projects with it — especially complex ones, where the cost of getting lost is highest and structure matters most.

03

A project others shape

More hands in its iteration. The organization itself is open — and meant to be argued with, extended, and improved.

04

Toward 150 roles

Expanding from 44 toward ~150 work roles across far more domains. As specialists arrive, it should grow from a software org into a general-purpose one.

05

Enjoy. Have fun.

That's the whole of it. Build something that matters, and enjoy being at your best while you do.

“You do not make beauty. Beauty emerges from things being true.”soul.md