CONCEPT Cited by 1 source
Three-layer agent platform stack¶
Definition¶
An emerging architectural decomposition for production AI agent platforms into three distinct responsibility layers:
| Layer | Responsibility | Examples |
|---|---|---|
| Framework | Project structure, conventions, integrations, CLI, developer experience | Flue |
| Harness | The agentic loop — tool calling, reading results, context management, continuation | Pi, Project Think, Codex, Claude Code |
| Runtime/Platform | Compute, state, and storage primitives everything above depends on | Agents SDK |
The key insight: a harness alone cannot solve production distributed-systems problems (crash recovery, secure sandboxing, durable state). These are platform concerns tied to storage and compute. Conversely, the platform shouldn't prescribe agent DX — that's the framework's job.
(Source: Agents platform post.)
Seen in¶
- sources/2026-06-17-cloudflare-bringing-more-agent-harnesses-and-frameworks-to-cloudflare — first explicit articulation of this three-layer decomposition
Related¶
- patterns/framework-harness-runtime-layering — the pattern this concept enables
- concepts/durable-execution — key runtime-layer primitive
- concepts/declarative-agent-definition — framework-layer design principle