Cognitecture manifesto

AI Leverage Matrix

A 2x2 feasibility screen for agentic work. Autonomy requires a separate warrant.

Two questions screen technical feasibility: Can the context be captured structurally? And is there a feedback loop to measure success?

Rich context plus fast feedback show technical feasibility, nothing more. An autonomy warrant sets what the agent may do, its limits, and its recourse.

Use the task lens carefully

The task lens aids analysis, not reassurance. AI acts on tasks. Firms and institutions decide how this changes jobs, pay, workload, and power.

The matrix screens technical feasibility. It cannot predict headcount, assign gains, or grant autonomy.

The feasibility screen

"Context is capturable"

Context includes sources, current state, instructions, and institutional memory. It is capturable when an agent can access and read what the task needs without a person rebuilding it each time.

Useful context needs provenance, an owner, update rules, proper access, conflict handling, and deletion. Some should stay uncaptured. Capturability helps feasibility. Permission and trust require more.

"Feedback loop"

A feedback loop gives the system evidence of success or failure, then uses it to improve the process. Signals include tests, metrics, rubric-based review, user behavior, or ground-truth checks.

A useful loop is fast, frequent, aligned, and actionable. A fast loop can still chase the wrong metric. Before relying on it, test its validity, independence, coverage, and resistance to manipulation.

What the matrix tells you

The matrix shows how easy candidate work is to produce and check. It does not decide trust or grant autonomy.

The matrix can change. Top-right is not every task's goal. Better context or feedback helps only when value exceeds assurance and absorption costs. Some context should stay uncaptured.

The grant decision: the autonomy warrant

The matrix screens feasibility. The autonomy warrant decides whether an agent may act without synchronous approval. Answer five questions:

Match autonomy to the consequence class: reversible, recoverable, hard to reverse, or safety-critical. A good demo cannot grant authority. The warrant is revocable; autonomy grows only with evidence. Read the full warrant in the manifesto →

Quadrant playbooks

Each quadrant suggests how to produce and check candidate work. The warrant still controls autonomy.

Human Leads. AI Assists.

Context: No / Feedback: No

This work depends on tacit knowledge, lived experience, or nuance. Success is hard to measure, reviewers often disagree, and confident output is hard to verify.

AI is useful for:

  • Generating options, not answers
  • Summarizing discussions and drafting material you will edit
  • Asking what could go wrong

Your job:

  • Set direction
  • Judge and own the result
  • Handle nuance and stakeholder dynamics

Watch out for:

Do not mistake fluent text for sound judgment. Values and tradeoffs need human decisions.

You're The Context Layer.

Context: No / Feedback: Yes

The loop works only when a person supplies context for every run. This is assisted execution; it grants no autonomous authority.

The workflow:

  • Human supplies a "Context Pack": goal, current state, constraints, examples, and changes
  • AI produces an output
  • The system returns feedback
  • Human updates context; AI tries again

Watch out for:

Watch for unstructured context drip, stale context, and AI chasing the metric instead of the goal.

Upgrade path:

Capture context from systems. Use intake forms and structured memory such as project KBs and decision logs. Turn repeated context into schemas, required fields, and retrieval sources.

You Judge. AI Executes.

Context: Yes / Feedback: No

AI can read the context and generate options, but success remains subjective. Judge each output before anyone acts.

The workflow:

  • Give AI retrieval, rules, and examples
  • Ask for multiple candidates (3-10)
  • Score them with a human rubric
  • Choose one and revise it
  • Record why

Watch out for:

"Looks good" without a rubric creates inconsistent results. Unrecorded feedback teaches nothing. Polished prose can hide false facts.

Upgrade path:

Label judgments and record reasons. Build QA checklists. Test proxy metrics against real outcomes. Structured judgment can improve assurance; proxies are not ground truth.

Feasible. Warrant Required.

Context: Yes / Feedback: Yes

Accessible context and measurable outcomes can make bounded agent execution feasible. An autonomy warrant must still authorize action.

Before action, write the warrant:

  • Define the delegated scope and excluded cases
  • Name the independent evidence and its limits
  • Grant least privilege with stop triggers
  • Size authority to the consequence class
  • Provide rollback, repair, recourse, and expiry

The production loop:

  • Ingest context (RAG / tools / structured data)
  • Generate with constraints
  • Validate (automated checks)
  • Execute only within the granted scope
  • Measure outcomes
  • Learn (update prompts, rules, retrieval, tests)

Watch out for:

Bad feedback rewards the wrong result. Watch for stale context and missing stop rules. Match autonomy to the consequence class; a demo cannot grant it.

Improve feasibility where it earns its cost

Better context or feedback can make candidate work easier. Invest only when value exceeds the cost of assurance, absorption, and context governance. Leave some tasks alone.

Move right: make context capturable

  • Use intake forms to capture explicit context
  • Standardize templates and schemas
  • Keep docs in one source of truth
  • Add good and bad examples plus edge cases
  • Build cited retrieval
  • Assign an owner and update schedule to each source

Move up: create a feedback loop

  • Define outcome measures and failure modes
  • Add validators: format checks, constraints, tests
  • Label human approvals and rejections
  • Run A/B tests where possible
  • Build a gold dataset of representative cases
  • Turn failures into tests through postmortems

Start with one task

Choose a frequent, repetitive, or costly task. Define its input, output, and failure conditions. Place it in the feasibility screen. If delegation fits, write the autonomy warrant before the agent acts.

The matrix does not decide jobs or human worth. It shows technical feasibility. People and institutions decide authority, rights, and who gets the gain.