| name | advanced-cognitive-architecture |
| description | Plan, reason, and verify across multiple levels using multi-step reasoning, tree-of-thought, and structured agent workflows. |
Skill: Advanced Cognitive Architecture
Category: Agent Intelligence
Priority: High
Description
This skill equips DevinOS with structured reasoning capabilities: multi-step planning, tree-of-thought exploration, assumption verification, and learning from mistakes. It helps the agent break complex problems into manageable reasoning chains and choose better actions.
Purpose
To make DevinOS a stronger problem solver by replacing shallow, single-step responses with deliberate, verifiable reasoning for complex engineering, design, and debugging tasks.
Trigger
Use this skill when:
- A task is ambiguous, multi-step, or has conflicting requirements.
- You need to compare multiple implementation approaches.
- A bug has several possible root causes and requires systematic elimination.
- You are planning a large feature or architectural change.
- A decision has significant downstream consequences.
Context
- The problem statement and constraints.
- Available tools, skills, and rules.
- Previous attempts and their outcomes (from memory).
- Risk tolerance and time budget.
- Verification criteria for success.
Workflow
- Decompose. Break the problem into sub-problems, constraints, and unknowns.
- Generate hypotheses. Produce multiple candidate approaches (Tree-of-Thought).
- Evaluate. Score each candidate against correctness, security, maintainability, performance, and cost.
- Select. Choose the best candidate or combine the best parts of several.
- Plan execution. Create a step-by-step plan with checkpoints and verification steps.
- Execute and verify. Run the plan and validate each step against the criteria.
- Reflect. If a step fails, backtrack, update hypotheses, and try again.
- Learn. Record what worked and what did not in memory.
Examples
Good: Choose a database architecture
User Input: "We need a real-time chat backend that supports 10k concurrent users. Should we use PostgreSQL, Redis, or a message queue?"
Agent Action:
- Decomposes: concurrency, persistence, ordering, scaling, failure modes.
- Generates candidates: PostgreSQL + LISTEN/NOTIFY, Redis pub/sub, Redis Streams, Kafka, RabbitMQ.
- Evaluates each against the requirements and trade-offs.
- Selects Redis Streams for persistence + scalability and a fallback message queue if needed.
- Plans implementation with migration and load-testing steps.
- Verifies and documents the decision.
Bad: Jump to the first idea without evaluation
User Input: "How do I handle 10k concurrent users?"
Agent Action:
- Refuses to give a one-size-fits-all answer.
- Asks about the domain, latency, durability, and budget.
- Applies the structured reasoning workflow before proposing a solution.
Anti-patterns
- Answering complex questions with a single shallow solution.
- Ignoring constraints and trade-offs.
- Failing to verify assumptions before building.
- Repeating the same failed approach without reflection.
- Producing long reasoning chains without a clear decision.
Verification
Cross Skill References
- AI Engineering: LLM prompt engineering, agent patterns, tool use.
- Planning: Task decomposition, sequencing, and risk assessment.
- Architecture: System design trade-offs and long-term consequences.
- Knowledge Distillation: Recording reusable reasoning patterns.
- Memory Manager: Persisting lessons from failures and successes.
References
