Skills vs MCP vs RAG vs Memory: a 4-part framework for what agents actually need to know
MaryamMiradi · x · 2026-09-16
AI researcher Maryam Miradi proposes a Procedure → Access → Knowledge → Experience framework for deciding what an agent should load into context during incidents—her core claim: more context doesn't make a better agent; each type of knowledge needs the right mechanism.
Step 1: Skills = Procedure
- Encode steps, rules, and judgment for repeatable tasks, pulled in only when needed
- Answer 'what should I do?', e.g., check error rate → inspect recent deployments → escalate if risk crosses a threshold
Step 2: MCP = Access
- Connects the agent to external systems and tools (truncated in original; the full version covers Knowledge and Experience layers)
She opens with the scenario of a production checkout returning 500 errors and flags the common anti-pattern of dumping runbooks, dashboards, docs, and customer history into context all at once.
More from coding & agent
- DeepSeek computer use draws a 'whale girl' in Pinta with mouse control, no MCP needed — teortaxesTex · 2026-09-16
- GitHub Copilot CLI v1.0.84-9 adds agent context management settings, ships batch of fixes — copilot-cli-release-app[bot] · 2026-09-16
- On-device speech transcription on Apple: SpeechAnalyzer and SpeechTranscriber resources — amos_gyamfi · 2026-09-16
- dspy-typesafeify: one decorator routes DSPy Signatures through Typesafe's typed inference — lateinteraction · 2026-09-16
- London Codex Meetup #4 returns Sept 21, 1,000+ past signups, OpenAI team presenting GPT-6 Astra — paw_lean · 2026-09-16
- OpenAI researcher: top staff run Codex agents 70+ hours a day, may upend pyramid orgs — daveholtz · 2026-09-16