Synapse Framework

Build AI teams
that actually ship work.

Synapse is the operating system for multi-agent teams — twelve teams, fifty-five active core agents, continuously audited (16,079 tool-call audit entries over 35 days). Distilled from long-running multi-agent practice.

12 Teams · 55 Agents · 16,079 audit entries · 29 D-records

Harness Engineering — Core Principle

Agent = Model + Harness

Models are a commodity. The harness is where the moat lives.

The Harness

Two control loops, one reliable team.

Guides shape what an agent does before it acts. Sensors verify what it produced after. Together they turn a raw model into a teammate you can rely on.

G

Guides — Before the act

Feedforward control. An agent knows its role, its rules, and where the guardrails sit before it writes a single line.

CLAUDE.md — role, rules, and workflow, version-controlled
Decision Rules — five escalation tiers
Agent Profiles — capabilities, grades, routing
Team Dispatch — mandatory routing table
S

Sensors — After the act

Feedback control. Nothing ships without evidence it passed.

QA Auto Review — four-dimension rubric, 85/100 to pass
Chain Audit — every step logged and checked
Expert Panel — multi-specialist cross-scoring
HR Audit Engine — quarterly capability review

Engagements

From diagnosis to build to ongoing evolution.

Three ways to work with us, each with a clear outcome and a published price.

Assessment

Synapse Assessment

A two-day readout that measures your current AI-team maturity against the Harness Engineering standard. Ends with a roadmap you own.

Contact · 2 Days

Pricing finalized in Phase 3 — see opc-strategy-2026-05-27

Implementation

Synapse Implementation

The full system, installed and running: execution chain, decision hierarchy, agent team, intelligence pipeline, HR audit.

Contact · 2-4 Weeks

Pricing finalized in Phase 3 — see opc-strategy-2026-05-27

Certification

SCP Training

Synapse Certified Practitioner. Eight hours of theory plus hands-on build. Attendees leave with a working AI-agent operating system — a repo, not slides.

Contact

Pricing finalized in Phase 3 — see opc-strategy-2026-05-27

Proof

Measured on a real team, not a demo.

Every number below came out of Synapse running in production — not a pitch deck.

synapse-dashboard — 12 teams · 55 agents · 35-day audit log
12
Teams
55
Agents
16,079
Audit Log · 35 days
29
Decisions Logged
CEO Guard audit slice
logs/ceo-guard-audit.log

The CEO Guard hook leaves a trace before and after every tool call. Below is a redacted slice of the audit log: timestamp, phase, tool type. Command contents and session identifiers are redacted — but the fact of the trace itself is governance evidence that cannot be forged.

2026-04-23 05:07PREBash
2026-04-23 05:08POSTBash
2026-04-23 05:09PREBash
2026-04-23 05:39PREEdit
2026-04-23 05:39POSTEdit
2026-04-23 05:40PREEdit
Displayable governance decisions (D-records)
title · date · level

Every governance-mechanism decision is archived under a number — the evidence trail from "incident" to "constitution."

L4
WF-09 exclusive Slack notification path
D-2026-05-19-002 · 2026-05-19
L4
Worktree → Main mandatory sync rule (P0 guard)
D-2026-0512-001 · 2026-05-12
L4
Weekly review mechanism + 4 governance structures
D-2026-0427-001 · 2026-04-27
L3
CLAUDE.md edit checklist — pre-commit hook enforced
D-2026-0427-003 · 2026-04-27
L3
Agent memory collaboration mechanism v1 enabled
D-2026-04-28-001 · 2026-04-28

Build in Public · Operations Story

When "success" is a lie

In early May 2026 we discovered something unsettling: Synapse's Slack notifications had been silently failing for over 11 days while the system reported "success" the entire time.

The root cause was a classic trap: the caller saw HTTP 200 and assumed success, while Slack's real verdict — {"ok":false,"error":"not_authed"} — sat inside the response body that nobody was reading. Worse still, the health heartbeat ran through the very same path, so the "monitor of the monitor" failed along with it. That is the monitoring paradox: when the sentry and the gate share one lock, a broken lock silences the sentry too.

Why did 11 days pass unnoticed? Because every link in the pipeline glowed green. GHA run status: green. HTTP status code: 200. At the technical layer "everything was fine" — at the business layer, nothing was being delivered. Technical success had masked business failure.

We did not treat it as a one-off bug to patch and forget. We turned the incident into institution:

  • A P0 governance principle was written into the system's constitution — the Silent Fail Defense Principle: "HTTP 2xx / status=success ≠ confirmed business effect; every external service call must assert on response-body business fields (e.g. ok=true); assertion failure must throw and mark the upstream job failed; silent stdout warnings are prohibited; meta-monitoring must be independent of the monitored pipeline." (Source: CLAUDE.md, tagged [ADDED: 2026-05-05])
  • Two weeks later we tightened it with a companion rule — the WF-09 Exclusive Notification Path: all Slack notifications must route through a single governed workflow; any bypass is a violation. (Source: CLAUDE.md [ADDED: 2026-05-19]; decision D-2026-05-19-002)
  • The single exception is a meta-monitoring channel independent of the main pipeline — built precisely to carry an alert out by another road when the main path fails as a whole. (Source: an independent meta-monitoring script, authorized exemption)

This is what we mean by "the Harness is not a claim, it is a trace": what you see is not a slogan but a dated rule, a numbered decision, and an auditable log line. The incident was not buried — it was recorded, institutionalized, and it permanently changed how the system behaves.

"Incidents become constitution — one WF-09 silent-fail incident becomes one P0 rule, one audit-log entry, one D-numbered decision. The harness isn't a claim; it's the trail."

— Synapse in-house operations data

Automation

The loop runs on a schedule — not in place of you.

Three daily windows + a weekly HR audit; L3 items dispatch autonomously, L4 items notify the Synapse Owner, L5 items pause for an Owner call.

06:00

Task Recovery

Scans unfinished work,
resumes what unblocked

08:00

Intelligence

Sweeps the frontier,
filters for signal

10:00

Action Pipeline

Triage → 4-expert panel
→ approved items ship

MON

HR Audit

Weekly capability review,
auto-remediated

Ready to build your AI team?

Start with the two-day assessment. You'll walk away with a measured baseline and a roadmap — regardless of what you do next.