Isagawa

Spec-driven loop engineering for AI agents.

Isagawa turns repeatable workflows into governed agent harnesses. The kernel enforces execution, the domain-spec factory builds vertical packs, and the backlog pipeline turns intent into validated work.

~/isagawa
See kernel OR SCROLL ↓

Why This Isn't Prompt Engineering

Execution Pipeline

Conversation / Intent
Backlog
Execute Pipeline
Task Builder
Kernel-Governed Agent
Contracts + Gates + Lessons
Validated Artifact / New Capability

Each step enforces contracts, validates gates, and records lessons for the next run.

Framework Architecture

Kernel
↓ runs
Domain Setup
↓ compiles
Domain Spec + Repo Context
↓ into
Protocol + Commands + Gates + Lessons
↓ enabling
Governed Agent Execution

Not configuration. Not templates. Specification-driven agent harnesses that learn from failures.

01

The Seed

The foundation: governance/runtime loop that all subsequent capabilities inherit. Hooks. Commands. Protocol. The framework that runs domain specs.

The kernel is a set of interlocking mechanisms, engineered through iterative specification and testing. Together they make governance a property of the system, not a layer on top of it. Every action the factory takes passes through these gates. The kernel blocks normal agent tool use when required protocol, validation, or learning steps are incomplete. Not optional. Not negotiable.

GOVERNANCE / PROTOCOL / TOKEN

Anchor Token

Every N actions (configurable), the system forces a full protocol re-read. A UUID token proves the agent actually re-centered. It cannot fake compliance by flipping a boolean.

HOOK / RUNTIME / WRITE-GATE

Gate Enforcer

A hook at the tool-call boundary that blocks writes until prerequisites are met. Missing a lesson? Blocked. Skipped an anchor? Blocked. The agent must fix the root cause to proceed.

LESSONS / MECHANICAL / PERSISTENCE

Learn Loop

Every failure records a lesson. The protocol updates mechanically. The same mistake becomes impossible. Not because the agent remembers, but because the system enforces.

STATE / LIFECYCLE / RECOVERY

Session Protocol

Start, anchor, work, complete. The same loop, every session. State survives restarts. Context survives compression. The system picks up where it left off.

Framework Architecture

Kernel (governance/runtime loop)

↓ runs

Domain Setup (bootstrapping/self-build loop)

↓ compiles

Domain Spec + Repo Context (installable vertical plugins)

↓ into

Protocol + Commands + Gates + Lessons

↓ enabling

Governed Agent Execution

Read the kernel landing page ->

02

Growth

SDD architecture: the kernel governs execution, the domain-spec factory compiles specifications into harnesses, and together they bootstrap each new capability.

The kernel governs execution. The domain-spec factory compiles specifications into structured agents. The result: 30+ AI agents, a 12-step compilation pipeline, and governed development environments that inherit kernel management from creation.

VERTICALS / CORPUS / SPECIFICATION

30+

AI Agents

Harnesses, agents, and domain specs compiled from specifications. Each one teaches the system a new field: QA, compliance, healthcare, DevOps, creative production, real estate.

PIPELINE / COMPILER / MECHANICAL

12

Agent Harness Factory Steps

A pipeline that compiles specifications into structured harnesses. The factory itself was produced by the kernel. The system built its own compiler.

ENVIRONMENTS / GOVERNANCE / INHERITANCE

GOVERNED

Workspaces

Complete development environments with hooks, commands, and protocols. Each workspace inherits kernel governance. Each one can produce new capabilities the kernel has never seen.

03

Self-Extension

The system produces new capabilities. Each capability extends what's possible next.

New skills, new commands, new workflows are produced and integrated. The system expands its own capabilities. Each new capability becomes part of the system, enabling even more capabilities.

SPECIFICATIONS / CAPTURE / EXECUTABLE

130+

/kernel/backlog

Every capability begins as a specification. The backlog captures specifications and structures them for execution. 130+ items logged, each one a starting point for autonomous production.

DECOMPOSITION / EXECUTION / AUTONOMOUS

90+

/kernel/execute-pipeline

One command decomposes intent into tasks and executes them. 90+ completed pipelines across 50+ repos. Autonomous for deterministic execution; HITL for approvals, failures, and judgment points.

OUTPUT / CAPABILITIES / RECURSIVE

Produced Capabilities

Produced Harnesses

Fraud Detection Platform · Healthcare QA · SSH Compliance Testing

Produced Skills

Website Cloner · Attestation Pipeline · Production Test Framework · Autonomous Cycling

and more. Each one produced by the same loop.

04

This Page

You are looking at the output.

This page is built by the spec-driven architecture it describes. The kernel provided governance. The backlog captured specifications. The pipeline decomposed and built. The attestation signed the work. Verifiable via Rekor timestamps below.

  1. Kernel engineered through specification
  2. Agent Harness Factory produced by kernel
  3. Agent Harnesses produced by factory
  4. Workspaces produced by harnesses
  5. /kernel/backlog specifications captured
  6. /kernel/execute-pipeline decomposed and built
  7. Signed to Rekor cryptographic provenance
  8. This Page the output

The loop is the core runtime architecture.

Provenance

Every pipeline run is signed with Sigstore and logged to Rekor. These are real attestation bundles. Verify them yourself.

View on Rekor ↗

        

View on Rekor ↗

        

View on Rekor ↗