Factual Knowledge
Project structure, module boundaries, API contracts, configuration conventions.
The Three Pillars of ANP
The Agent-Native Programming paradigm is built on three foundational pillars. Each addresses a fundamental requirement for agents to operate as first-class entities in software development.
Pillar 1: Agent-Computer Interface (ACI)
Key Principle: Interfaces should encode knowledge about correct usage, reducing cognitive burden on agents while enabling structural safety guarantees.
The ACI provides structured protocols for agent-to-system communication. Unlike raw shell access or unstructured APIs, the ACI is designed so that correct usage is easy and misuse is structurally difficult.
Key Aspects
- Semantic operations — Expose intent, not mechanics (
edit_fileinstead ofwrite_bytes) - Discoverable affordances — Agents query what’s available and learn constraints
- Structural guardrails — Safety is embedded in the interface, not delegated to agent discipline
- Composability — Operations chain, parallelize, and nest cleanly
- Graceful degradation — Structured errors with corrective suggestions
Protocol Layers
┌─────────────────────────────────────────────┐│ Layer 4: Semantic Layer ││ (Domain-specific operations and workflows) │├─────────────────────────────────────────────┤│ Layer 3: Capability Layer ││ (Tool definitions, permissions, discovery) │├─────────────────────────────────────────────┤│ Layer 2: Session Layer ││ (Context management, state, transactions) │├─────────────────────────────────────────────┤│ Layer 1: Transport Layer ││ (Message encoding, delivery, framing) │└─────────────────────────────────────────────┘ Full ACI Specification Deep dive into the protocol architecture, message taxonomy, and implementation patterns.
Pillar 2: Semantic Memory
Key Principle: Memory is not just storage — it is an active, evolving system that makes an agent more valuable in its hundredth session than its first.
Drawing from cognitive science (Atkinson-Shiffrin, Tulving, Baddeley), ANP proposes a three-layer memory architecture:
Memory Layers
┌─────────────────────────────────────────────┐│ Working Memory ││ Current session context, active task state ││ Capacity: Limited | Lifetime: Session │├─────────────────────────────────────────────┤│ Semantic Memory ││ Facts, patterns, relationships, skills ││ Capacity: Large | Lifetime: Long-term │├─────────────────────────────────────────────┤│ Episodic Memory ││ Events, interactions, decisions, outcomes ││ Capacity: Large | Lifetime: Long-term │└─────────────────────────────────────────────┘What Agents Remember
Procedural Knowledge
Build processes, deployment procedures, review conventions, release workflows.
Episodic Knowledge
Design decisions, rejected approaches, fragile code areas, team preferences.
Meta-Knowledge
Self-calibration data, learned heuristics, effective tools and approaches.
Full Memory Architecture Complete analysis of memory requirements, lifecycle management, and cross-agent sharing.
Pillar 3: Natural Language as Lingua Franca
Key Principle: Natural language is not just for prompts — it’s the architectural interface for task specification, progress reporting, and collaboration.
In ANP, natural language serves as the primary medium of communication between:
- Human to Agent — Task specification, feedback, oversight
- Agent to Human — Progress reporting, clarification requests, decision proposals
- Agent to Agent — Coordination, negotiation, knowledge sharing
Why Natural Language?
- Expressiveness — Captures nuance, intent, uncertainty, and context that typed interfaces cannot
- Accessibility — No special tooling needed for human participants
- Flexibility — Adapts to novel situations without schema changes
- Partnership — Positions agents as colleagues, not just tools
This doesn’t mean abandoning structure — the ACI provides structured operations for precise actions. Natural language handles the spaces between: planning, reasoning, explaining, and collaborating.
Multi-Agent Coordination How agents communicate and coordinate using natural language protocols.