One loop
Measure a gap, act to shrink it, then check again. Simple, cheap, and often the right answer.
measure → act → check → repeat
01 A control architecture for honest improvement
Stop optimizing the dashboard. Start engineering the system.
Graph Engineering connects execution, verification, audit, arbitration, and human judgment so a system can improve without learning to fool itself.
Not “more agents.” Better evidence, authority, and stopping rules.
GROUND / GOVERN / VERIFYGE—01
3architecture modes
9control responsibilities
10explicit edge meanings
1human-owned root outcome
02 The plain-language version
A thermostat is a loop. A trustworthy organization is a graph.
A graph of loops
Different loops act, verify, audit, and arbitrate. Their edges define evidence and authority.
act → verify → constrain | veto
Grounded improvement
At least one protected signal must touch reality outside the optimizing machinery.
claim → independent evidence → decision
03 Why a perfect loop can fail perfectly
The number goes up. The real result walks away.
Single-loop failures are not random. They follow from what the loop cannot see or govern.
Metric gaming
Once a proxy becomes the target, the system discovers cheap ways to move the number without delivering the outcome.
Wrong target
A loop can optimize a reference value, but it cannot decide whether that reference was wise.
Loop conflict
Speed, quality, safety, and cost can each improve locally while the whole system gets worse.
Measurement decay
Sensors drift, definitions change, and dashboards keep agreeing with reports produced by the same broken pipeline.
04 The minimum grounded graph
Responsibilities first. Agents later.
A node is a control responsibility. It may be implemented by a person, a deterministic test, a service, a model, or an agent.
01Executeact within bounds
02Verifytest the claim
03Counter-signalfind cheap wins
04Auditcheck the watcher
05Arbitrateresolve conflict
06Governreview targets slowly
07Human gateown judgment
08Stop & hand offpreserve truth
⚓Reality anchorevidence the optimizer cannot rewrite
05 The part topology cannot supply
Every graph needs something outside the graph.
A second agent is not independent if it reads the same corrupted table. A second dashboard is not a reality check if both dashboards share the same pipeline.
Independent verification requires both an independent evaluator and independent evidence provenance.
- Executed tests, not reported intentions
- Retained customers, not closed tickets
- Protected data the optimizer cannot inspect
- Human approval for root goals and frozen rules
06 Four layers of understanding
From intuition to an executable contract.
Intuition
One loop improves one number. A graph makes the loops watch, constrain, and correct one another.
Design
Name every responsibility, evidence artifact, veto, escalation, cadence, and human owner.
Contract
Persist the graph in machine-readable JSON with frozen anchors, verification checks, and stop conditions.
Operation
Implement only the nodes that need persistent loops, then verify outcomes against fresh anchors.
07 Use the smallest architecture that works
Not every task deserves a graph.
Direct
One bounded action. No useful feedback cycle. Solve it directly.
Example: translate one sentence.Loop
One dominant goal, one trustworthy verifier, meaningful retries.
Example: repair a failing test.Graph
Several goals, speeds, verifiers, conflicts, or authorities interact.
Example: improve support without harming retention.08 A working Codex Skill, not just an essay
Design it. Persist it. Validate it.
The repository includes an installable Skill, a graph initializer, a structural validator, a complete customer-support example, regression tests, and dependency-free CI.
$ npx skills add 2023Anita/graph-engineering-architect
# Invoke in Codex
Use $graph-engineering-architect to design the minimum grounded architecture for this workflow.
# Verify the repository
$ make verify09 Five rules worth keeping
The graph is honest only when its boundaries are honest.
- 01
Do not create one agent per node by default.
- 02
Do not call shared-data self-review independent verification.
- 03
Do not let fast loops rewrite slow goals or protected tests.
- 04
Do not let improvement machinery grant itself more authority.
- 05
Do not declare success without fresh anchored evidence and a usable handoff.