NOTE

Cognitive Architectures

authorgemini-cli aliaseschain-of-thought-blueprint, agentic-workflows, project-lifecycle-patterns titleCognitive Architectures statusactive date2026-04-25 typepermanent

Cognitive Architectures

A Cognitive Architecture for AI agents provides a structured map of how to "think" through a complex project. It prevents agents from getting stuck in Local Minima (fixing symptoms while missing the disease) by enforcing a clear lifecycle from discovery to verification.

The Bespoke Software Lifecycle

To build high-quality custom software, agents should follow this standard cognitive flow:

1. Research (Discovery Phase)

  • Codebase Mapping: Systematically exploring files and symbols.
  • Dependency Analysis: Identifying how components interact.
  • Empirical Reproduction: Confirming the current state or bug before changing anything.
  • Scope Confirmation (Gate): Explicitly listing what is out of scope to prevent intent drift.
  • Output: A Scope Confirmation block defining the task perimeter.

2. Strategy (Design Phase)

  • Architectural Alignment: Ensuring the solution fits existing patterns (e.g., YANP).
  • Tool Selection: Choosing the most efficient MCP tools or libraries.
  • Plan Formulation: Drafting a step-by-step implementation guide.
  • Safety & Boundaries Checkpoint: A mandatory pre-flight check before Execution:
      1. Are all planned write targets within authorized scope?
      2. Does the plan include irreversible operations (delete, push, overwrite uncommitted changes)?
      3. Is there a human-review requirement for this action class?
      4. Does any planned action touch secrets, credentials, or PII?
  • Output: A written Implementation Plan (even a 3-bullet list).

3. Execution (The Act-Validate Loop)

  • Surgical Implementation: Applying targeted changes strictly related to the sub-task.
  • Automated Verification: Running tests or linters immediately after changes.
  • Refinement: Adjusting the approach based on system feedback.
  • Loop Termination: If verification fails after 3 iterations, the agent must stop, record the failure state, and escalate to human review.
  • Output: A Passing Test Run or a detailed Failure Report.

4. Finalization (System Integrity)

  • Cross-Linkage: Integrating new notes or code into the broader graph.
  • Regression Check: Running broken-link and orphan scanners to ensure no side-effect errors were introduced.
  • Portal Update: Synchronizing dashboards and indices.
  • Output: A green run-maintenance.ps1 result.

Agentic Workflow Patterns (Andrew Ng)

Frontier models leverage these four patterns to increase performance:

  1. Reflection: The agent critiquing its own output to fix hallucinations or logic gaps.
  2. Planning: Breaking a high-level goal into a sequence of executable steps.
  3. Tool Use: Leveraging external capabilities (calculators, web search, database queries).
  4. Multi-Agent Collaboration: Dividing labor among specialized personas (e.g., Architect, Coder, Judge).

Framework Implementation: LangGraph

Frameworks like langgraph enable these architectures by representing the lifecycle as a Stateful Graph. This ensures that the agent's "shared memory" (State) persists across the entire project lifecycle, allowing for loops, branches, and Human-in-the-Loop checkpoints.

References