How It Works

Agent = Model + Harness.

Everyone has access to the same models. The harness is where reliability is engineered. Synapse is a reference implementation of Harness Engineering — the 2026 engineering discipline for production AI teams.

The Mechanics

Guides and Sensors

A harness is two things working together. Guides are feed-forward: they shape behavior before action — the CLAUDE.md persona, the organization.yaml routing table, the decision_rules.yaml constraints, the agent capability cards.

Sensors are feedback: they watch behavior after action — the 85/100 QA review, the execution-chain auditor, the decision log, the cross-session state file.

Guides stop drift. Sensors catch it when it happens anyway.

The Process

The execution chain

Five steps, every time. Violating any one of them is logged as a P0 break.

1

Greeting + objective intake.

Lysander confirms identity and restates the goal.

2

Classification.

The execution_auditor grades the task S/M/L. S ships fast. L goes through expert review.

3

Mandatory dispatch table.

No work starts without a named specialist and a named deliverable. This is a hard constraint.

4

Isolated execution.

Specialists run in sub-agent contexts. The CEO persona never touches Edit/Write/Bash directly — CEO Guard blocks it.

5

QA gate + delivery.

integration_qa scores the output on integrity, accuracy, consistency, maintainability, compliance. ≥85 ships.

The Cache Architecture

The three-layer CLAUDE.md

Synapse's CLAUDE.md is three layers with distinct cache lifetimes.

  • Layer 1 Identity (~25 lines) — changes almost never.
  • Layer 2 Harness rules (~170 lines) — changes on upgrades.
  • Layer 3 Instance config (~85 lines) — changes on personalization.

The ordering is deliberate: stable content first so the Anthropic prompt cache stays valid across turns, cutting cost by ~90% on hits.

The Governance

Entropy budget

Every rule in Synapse carries a timestamp. Every 3 days, harness_engineer runs audit_harness() against the CLAUDE.md file — line count, duplicate semantics, stale rules past 180 days.

Above 300 lines, new additions are blocked until something is removed. Governance is a first-class feature, not an afterthought.

Ready to run it yourself?

Download the configuration bundle. Clone, swap three variables in CLAUDE.md, restart your Claude Code session — and the first dispatch can run.