The cognitive loop

The cognitive loop is the core abstraction that makes agent reasoning explicit and auditable. Every Noēsis episode flows through a sequence of nine phases.

The nine phases

The governance phase acts as a critical gate—if policies veto the plan, execution jumps to a blocked state. Each phase has a specific purpose and produces structured events that form the episode timeline.

In minimal mode, the Direction, Governance, and Insight phases are skipped for faster execution. In meta mode (default), all nine phases execute with full observability.

Observe

The observe phase captures the raw input at the moment an episode starts. Purpose: Record exactly what the agent was asked to do, with all context. What gets recorded:

Task text (the goal or prompt)
Tags (metadata like environment, priority)
Timestamp
Initial context

Example event:

{
  "phase": "observe",
  "payload": {
    "task": "Draft release notes for v1.2.0",
    "tags": {"priority": "high", "team": "platform"},
    "timestamp": "2024-01-15T10:30:00Z"
  },
  "metrics": {
    "duration_ms": 1
  }
}

Why it matters: You can confirm the exact scope the agent perceived, enabling accurate replay and debugging.

Interpret

The interpret phase extracts signals and intent from the observed input. Purpose: Summarize what the policy or intuition layer noticed before any plan is locked in. What gets recorded:

Signals (risks, opportunities, constraints)
Intent classification
Relevant context from memory
Policy observations

Example event:

{
  "phase": "interpret",
  "payload": {
    "signals": [
      {"type": "risk", "description": "Production deployment mentioned"},
      {"type": "constraint", "description": "Requires approval for releases"}
    ],
    "intent": "documentation_generation"
  },
  "caused_by": "observe_event_id",
  "metrics": {
    "duration_ms": 5
  }
}

Why it matters: You can see what influenced planning decisions, making the reasoning chain transparent.

Plan

The plan phase decides what actions to take. Purpose: Record the selected steps so you can compare intent versus action. What gets recorded:

Ordered steps with descriptions
Tools or adapters to invoke
Expected outcomes
Confidence scores

Example event:

{
  "phase": "plan",
  "payload": {
    "steps": [
      {"kind": "detect", "description": "Gather changelog entries", "status": "pending"},
      {"kind": "analyze", "description": "Categorize changes", "status": "pending"},
      {"kind": "act", "description": "Generate release notes", "status": "pending"},
      {"kind": "verify", "description": "Check formatting", "status": "pending"}
    ],
    "confidence": 0.85
  },
  "caused_by": "interpret_event_id",
  "metrics": {
    "duration_ms": 12
  }
}

Step kinds: The plan uses a controlled vocabulary for step types:

Kind	Purpose
`detect`	Gather information
`analyze`	Process or categorize
`plan`	Sub-planning
`act`	Execute action
`verify`	Check results
`review`	Human review point

Why it matters: You can audit what was planned and detect drift from the original intent.

Act

The act phase executes the planned actions. Purpose: Log every tool or adapter invocation with inputs and outcomes. What gets recorded:

Tool or adapter name
Input excerpt
Output or result
Execution metrics

Example event:

{
  "phase": "act",
  "payload": {
    "tool": "changelog_reader",
    "input_excerpt": "CHANGELOG.md",
    "outcome": {"entries_found": 15},
    "status": "success"
  },
  "caused_by": "plan_event_id",
  "metrics": {
    "started_at": "2024-01-15T10:30:01Z",
    "completed_at": "2024-01-15T10:30:02Z",
    "duration_ms": 1200
  }
}

Why it matters: You get a measurable execution history instead of guesswork about what happened.

Reflect

The reflect phase evaluates what actually happened. Purpose: Compare outcomes against expectations and record the assessment. What gets recorded:

Success/failure status
Reasons for the outcome
Comparison to expected results
Issues encountered

Example event:

{
  "phase": "reflect",
  "payload": {
    "success": true,
    "reason": "All steps completed successfully",
    "expected_outcomes": ["release_notes_generated", "format_verified"],
    "actual_outcomes": ["release_notes_generated", "format_verified"],
    "issues": []
  },
  "caused_by": "act_event_id",
  "metrics": {
    "duration_ms": 3
  }
}

Why it matters: Dashboards can alert on failures, and you can analyze patterns in successes and failures.

Learn

The learn phase captures updates for future runs. Purpose: Record follow-up proposals so the next run can inherit lessons. What gets recorded:

Update proposals
Scope of changes
Memory updates
Policy adjustment suggestions

Example event:

{
  "phase": "learn",
  "payload": {
    "updates": [
      {"type": "memory", "key": "changelog_format", "value": "keep-a-changelog"},
      {"type": "hint", "content": "Include breaking changes section"}
    ],
    "scope": "session"
  },
  "caused_by": "reflect_event_id",
  "metrics": {
    "duration_ms": 2
  }
}

Why it matters: Episodes can improve over time without manual intervention.

Insight

The insight phase computes KPIs and metrics from the episode. Purpose: Generate structured metrics for dashboards, alerts, and analysis. What gets computed:

Plan adherence (how closely execution matched the plan)
Veto count
Tool coverage
Latency percentiles
Custom KPIs

Example event:

{
  "phase": "insight",
  "payload": {
    "metrics": {
      "plan_adherence": 0.95,
      "veto_count": 0,
      "tool_coverage": 1.0,
      "branching_factor": 2
    },
    "kpis": {
      "success_rate": 1.0,
      "first_action_latency_ms": 150
    }
  },
  "caused_by": "learn_event_id"
}

Why it matters: Structured KPIs enable automated monitoring, alerting, and continuous improvement.

Phase instrumentation

Since v0.7.0, every phase is instrumented with timing and lineage:

{
  "id": "7d3d...f84",
  "phase": "plan",
  "payload": {...},
  "metrics": {
    "started_at": "2024-01-15T10:30:00.500Z",
    "completed_at": "2024-01-15T10:30:00.512Z",
    "duration_ms": 12.7
  },
  "caused_by": "5c12...0a9"
}

started_at / completed_at: High-resolution timestamps
duration_ms: Phase execution time
caused_by: UUID linking to the causal parent event

This enables:

Performance profiling per phase
Causal chain reconstruction
Bottleneck identification

Direction

The direction phase applies policy-driven plan mutations (meta mode only). Purpose: Allow policies to modify the plan before execution based on intuition signals. What gets recorded:

Directive ID (deterministic UUIDv5 for lineage)
Status (applied, blocked, skipped)
Diffs showing what changed
Policy information

Example event:

{
  "phase": "direction",
  "payload": {
    "directive_id": "dir_abc123",
    "status": "applied",
    "advice": "Added safety bounds to query",
    "diff": ["plan.steps[0].parameters"],
    "policy_id": "SafetyPolicy@1.0",
    "confidence": 0.85
  },
  "caused_by": "plan_event_id"
}

Why it matters: You can see exactly how policies modified the plan before execution.

Governance

The governance phase is a critical gate that audits actions before execution (meta mode only). Purpose: Enforce pre-action policies and provide audit trails for compliance. What gets recorded:

Governance ID (deterministic UUIDv5)
Decision (allow, audit, veto)
Rule that triggered the decision
Confidence score

Example event:

{
  "phase": "governance",
  "payload": {
    "governance_id": "gov_def456",
    "decision": "allow",
    "rule_id": "rules.allow.default",
    "score": 0.95,
    "policy_id": "governance.rules",
    "policy_version": "1.0.0"
  },
  "caused_by": "direction_event_id"
}

Governance decisions:

Decision	Effect
`allow`	Action proceeds normally
`audit`	Action proceeds, flagged for review
`veto`	Action blocked, episode ends

Why it matters: The PreActGovernor ensures dangerous actions are blocked before they execute, with full audit trails.

Event order invariant

Events follow this order in meta mode:

observe → interpret → plan → direction → governance → act+ → reflect → learn → insight → terminate

In minimal mode, direction, governance, and insight are skipped:

observe → interpret → plan → act+ → reflect → learn? → terminate

Invariant: Even on an error or veto, Noēsis emits an ordered trace and summary. You always get artifacts.

This means:

Failed episodes still have complete timelines
Vetoed episodes record why they were blocked (governance decision recorded)
Errors are captured in the reflect phase
You can always inspect what happened

Feedback loops

The diagram shows two feedback paths:

Insight → Interpret: Reflections from one cycle can inform the next interpretation
Learn → Observe: Adaptations can modify how future observations are processed

These enable:

Progressive refinement within an episode
Cross-episode learning
Policy adaptation over time

Human in the loop

Human review slots naturally between Act and Reflect: Policies can flag operations for human approval, pausing the loop until a decision is made.

Reading the timeline

Use the CLI to inspect the timeline:

# All events
noesis events ep_abc123

# Filter by phase
noesis events ep_abc123 --phase plan

# As JSON for scripting
noesis events ep_abc123 -j | jq '.[] | select(.phase == "act")'

Or in Python:

import noesis as ns

events = list(ns.events.read(episode_id))

# Filter phases
plan_events = [e for e in events if e["phase"] == "plan"]
act_events = [e for e in events if e["phase"] == "act"]

# Reconstruct causal chain
def get_causal_chain(events, event_id):
    chain = []
    current = next((e for e in events if e["id"] == event_id), None)
    while current:
        chain.append(current)
        current = next((e for e in events if e["id"] == current.get("caused_by")), None)
    return chain

Get started

Tutorials

How-to guides

Explanation

The cognitive loop

The nine phases

Observe

Interpret

Plan

Act

Reflect

Learn

Insight

Phase instrumentation

Direction

Governance

Event order invariant

Feedback loops

Human in the loop

Reading the timeline

Next steps

Faculties

Events reference

Get started

Tutorials

How-to guides

Explanation

​The nine phases

​Observe

​Interpret

​Plan

​Act

​Reflect

​Learn

​Insight

​Phase instrumentation

​Direction

​Governance

​Event order invariant

​Feedback loops

​Human in the loop

​Reading the timeline

​Next steps

Faculties

Events reference

The nine phases

Observe

Interpret

Plan

Act

Reflect

Learn

Insight

Phase instrumentation

Direction

Governance

Event order invariant

Feedback loops

Human in the loop

Reading the timeline

Next steps