control-mirror

Audits agent and system architecture through a control theory lens for stability, feedback, noise, delay, oscillation, error control, adaptive behavior, and...

Control Mirror turns the core ideas of Qian Xuesen's Engineering Cybernetics into a practical architecture review skill.

It is not a generic architecture review checklist. It treats a system as a controlled dynamic system and asks one brutal question:

Is this system becoming more controllable, stable, observable, and adaptive — or merely more complicated?

Use it as three tools at once:

• Mirror: reveal hidden instability, noise, delay, positive feedback, and uncontrolled loops.
• Ruler: measure whether the architecture has real controllability, not just many modules.
• Compass: identify the next highest-leverage evolution step toward a self-stabilizing system.

---

When to Use

Use this skill when the user wants to review, diagnose, or evolve any system architecture, especially:

• software platforms, agent systems, workflow engines, automation systems, data pipelines, product operating systems
• multi-agent systems, AgentOS, LLM tool-use systems, memory systems, routing layers, evaluation pipelines
• systems that repeatedly fail, oscillate, drift, over-consume resources, or require constant human supervision
• architectures with unclear feedback loops, weak observability, noisy memory/context, delayed correction, or unsafe automation
• reviews involving: control theory, feedback, stability, damping, adaptation, error correction, delay, noise, self-stabilization, cybernetics

Do not use this as a generic “is the code clean?” review. Use it when the key question is system behavior over time.

---

Core Lens

Do not start by asking “how many features does the system have?”

Start with:

1. What is the reference input / target state?
2. What acts as the controller?
3. What is the controlled object / environment?
4. What sensors observe the system state?
5. What feedback changes future behavior?
6. Where are delay, noise, saturation, and error accumulation introduced?
7. What prevents unstable positive feedback?
8. Can the system converge after disturbance?

If you cannot map these, the review is still too superficial.

---

Engineering Cybernetics Mapping

Map the user's architecture into this structure:

Cybernetics concept	Architecture equivalent
Reference input	User goal, SLA, product target, policy, task objective
Controller	Scheduler, orchestrator, agent, workflow engine, governance layer
Controlled object	Codebase, service, team process, model provider, external environment
Actuator	Tool calls, deployments, writes, API calls, workflow transitions
Sensor	Tests, logs, metrics, review, cost reports, human feedback, evals
Feedback	Signals that change later routing, budget, memory, workflow, or policy
Noise	Irrelevant context, stale memory, bad retrieval, flaky logs, misleading metrics
Delay	Async queues, slow tools, stale state sync, delayed human review
Error	Difference between target and actual state
Damping	Rate limits, compression, budget caps, confirmations, backoff, gates
Saturation	Token limits, API limits, budget limits, human attention limits
Stability boundary	Conditions where automation must pause, degrade, or ask for confirmation

---

Review Workflow

Phase 1: Define the Control Loop

Produce a concise control-loop map:

Target → Controller → Actuator → Controlled object → Sensor → Feedback → Controller

Then identify:

• feedback frequency
• feedback quality
• delay points
• noise sources
• constraints / saturation points
• human checkpoints

Phase 2: Detect Instability Patterns

Look for the highest-risk 3-5 patterns only. Do not dump a huge checklist.

1. Open-loop execution

Symptoms:

• system executes but does not verify
• failures do not change future strategy
• humans discover errors only after damage

Control diagnosis: the system lacks effective feedback.

2. Positive feedback amplification

Symptoms:

• bad memory causes worse retrieval, which writes more bad memory
• failing retries increase cost and context noise
• low-quality outputs feed later decisions without validation

Control diagnosis: error is amplified instead of damped.

3. Oscillation

Symptoms:

• repeated searching, repeated rewriting, repeated re-planning
• agents hand tasks back and forth
• workflow loops without new evidence

Control diagnosis: feedback exists, but damping and convergence criteria are weak.

4. Delay-induced correction failure

Symptoms:

• async feedback arrives after the system has already moved on
• state is stale when decisions are made
• human approval comes too late to prevent error propagation

Control diagnosis: the feedback loop has harmful latency.

5. Noise pollution

Symptoms:

• irrelevant context dominates current task signal
• stale memory is treated as truth
• logs are too verbose to reveal failure causes

Control diagnosis: the sensor layer lacks filtering.

6. Error accumulation

Symptoms:

• small deviations silently compound
• no checkpoint validates intermediate state
• rollback is missing or expensive

Control diagnosis: there is no bounded-error mechanism.

7. Fake adaptation

Symptoms:

• system looks flexible, but every adjustment is manual
• policies do not update from observed outcomes
• “AI decides” but cannot explain or revise its decision rule

Control diagnosis: adaptation is not closed-loop.

8. Saturation blindness

Symptoms:

• token/cost/API/human attention limits are treated as infinite
• system fails only after hitting hard limits
• no graceful degradation exists

Control diagnosis: constraints are not part of the controller.

Phase 3: Measure Real Strengths

Only count an architectural strength if it improves controllability.

Good strengths include:

• feedback that actually changes future behavior
• tests/reviews that gate irreversible actions
• cost/token/resource constraints that affect routing
• memory that is filtered, layered, and reversible
• observability that exposes why decisions were made
• fallback and degradation paths
• explicit pause/confirm conditions for high-risk actions

Do not praise “many modules”, “many agents”, or “complex workflows” unless they improve stability.

Phase 4: Score Control Maturity

Use this 0-5 scale:

Level	Name	Meaning
0	Open loop	Executes without reliable feedback
1	Human-corrected loop	Humans catch and correct most errors
2	Verified loop	Tests/reviews/metrics gate some actions
3	Damped closed loop	Budgets, gates, retries, and fallbacks reduce instability
4	Adaptive closed loop	Historical outcomes tune future routing/policy/budget
5	Self-evolving controlled system	The system improves its own procedures while preventing drift and pollution

Give a level with one sentence of evidence.

Phase 5: Recommend Evolution

Prioritize only 1-3 actions.

Use this priority logic:

• P0: Prevent loss of control — add gates, rollback, confirmation, error bounds, loop detection.
• P1: Improve self-stabilization — add damping, filtering, degradation, better observability.
• P2: Improve adaptation — use historical outcomes to tune policies, promote SOPs, evolve workflows.

---

Control Scorecard

When the user wants a rigorous review, include this table:

Dimension	Score / 5	What to check
Feedback completeness		Does output affect future decisions?
Stability / convergence		Does the system stop oscillating and converge?
Noise filtering		Are irrelevant/stale signals filtered before decisions?
Delay handling		Are stale/late signals detected and handled?
Error control		Are small errors bounded before they cascade?
Damping / resource control		Are cost/token/API/human limits part of control?
Observability		Can humans see why decisions were made?
Adaptation		Do historical outcomes tune future behavior?
Safety boundary		Does automation pause on irreversible/high-risk actions?

Then summarize:

Control maturity level: L0-L5
Strongest control loop: ...
Weakest control loop: ...
Highest-risk instability: ...
Next P0 action: ...

---

Output Format

Use this default structure:

1. Control-system mapping

One concise paragraph or diagram mapping the system into controller / controlled object / feedback / noise / delay / constraints.

2. Architecture problems — Mirror

List the most important 3-5 problems.

For each:

• Problem: what is unstable or uncontrolled
• Control diagnosis: why this is a feedback/noise/delay/error issue
• Consequence: what happens if it remains unfixed

3. Architecture strengths — Ruler

Only list strengths that improve controllability, stability, observability, or adaptation.

4. Evolution direction — Compass

Give 1-3 prioritized actions:

• priority
• action
• why it comes first
• how to verify it worked

5. Control scorecard

Include when the review is non-trivial or when the user asks for scoring.

---

Domain-Specific Review Prompts

Software platform / microservices

Check:

• service health signals → routing / scaling / rollback
• alert delay and alert fatigue
• retry storms and cascading failures
• circuit breakers and backpressure
• deployment rollback and blast-radius control

Data pipeline / analytics system

Check:

• data quality sensors
• late-arriving data handling
• schema drift detection
• bad data quarantine
• downstream contamination recovery

Workflow / operations system

Check:

• whether each step has a gate
• whether failed steps loop back to the right point
• whether work-in-progress creates hidden queues
• whether handoffs introduce delay or noise
• whether “done” has measurable criteria

Agent / LLM system

Check:

• tool calls are verified, not assumed
• memory is layered and filtered
• retrieval errors do not poison future context
• model routing considers complexity, cost, latency, health, and outcome history
• long outputs are compressed before entering context
• high-risk writes/actions require confirmation

Organization / socio-technical system

Check:

• decision feedback is timely enough to matter
• local team incentives do not amplify global instability
• metrics do not become noisy proxies
• governance prevents irreversible mistakes without freezing execution

---

AgentOS / Multi-Agent Extension

Use this section only when reviewing AgentOS, multi-agent platforms, OpenClaw-like systems, LLM orchestration, or autonomous workflow engines.

AgentOS control-loop map

User goal / task queue
  → execution kernel / orchestrator
  → complexity scoring + routing policy
  → model / agent / tool execution
  → tests / review / cost / logs / user feedback
  → memory / metrics / SOP candidates
  → next routing, workflow, budget, or policy decision

AgentOS-specific questions

• Is there a single default execution entrypoint, or can components bypass the control loop?
• Does model routing use complexity, budget, health, latency, historical success/failure, and user preference?
• Are high-token outputs such as test logs, build logs, git diffs, search/list/tree outputs compressed and measured?
• Do workflow templates have gates and failure loops, or are they only diagrams?
• Does memory separate runtime observations, task summaries, failure cases, SOP candidates, and long-term wiki notes?
• Do failures and costs tune future behavior, or are they only recorded?
• Are irreversible writes, public actions, deletions, deployments, and over-budget actions bounded by confirmation?

AgentOS scorecard add-on

Dimension	Score / 5	Deduct when...
Default entrypoint unity		APIs/tools bypass the kernel/orchestrator
Model routing adaptiveness		routing is keyword-only or ignores history/cost/health
Token damping		noisy outputs enter context uncompressed
Workflow gate strength		steps can advance without required evidence
Memory pollution resistance		temporary noise is written into long-term memory
Feedback-to-policy loop		failures/costs are logged but do not affect future decisions
Explainability		decisions lack request id, factors, or human-readable reason
Safety boundary		high-risk automation has no pause/confirm path

Common AgentOS instability patterns:

• Kernel bypass: multiple execution paths ignore the main controller.
• Feedback as logs only: sensors exist, but do not affect control.
• Memory writeback overheating: every observation becomes long-term truth.
• Formal workflow gates: steps exist, but missing evidence does not block progress.
• Unobservable damping: compression/budgeting exists, but savings and loss are not measurable.

---

Anti-Patterns

Avoid these mistakes:

• Turning the review into a theoretical control theory lecture.
• Praising complexity as sophistication.
• Listing every possible issue instead of the few that most affect control.
• Suggesting adaptation before the system has basic stability.
• Ignoring human attention and budget as hard constraints.
• Treating logs as feedback when they do not change future decisions.
• Treating memory as knowledge without filtering, promotion, or rollback.

---

Quality Bar

A good Control Mirror review must be:

• Mapped: clear controller / object / sensor / feedback structure.
• Diagnostic: names the control failure, not just the symptom.
• Prioritized: gives P0/P1/P2, not a flat wishlist.
• Verifiable: each recommendation has a way to check completion.
• Grounded: tied to the actual architecture, code, workflow, or process.

If the output does not help the user make the system more controllable, stable, observable, or adaptive, it failed.