Coding Agent as Foundation for General-Purpose Agents¶
Central argument of harrison-chase and sam-partee:
"Coding agents are actually the foundation for general purpose agents. This might not immediately make sense, but hopefully over the course of the talk it does make more sense."
The move that coding agents made — file-system memory, tool-mediated retrieval, iterative planning — generalises to any domain agent. You don't need new primitives; you need the same primitives with different tools attached.
What coding agents already do¶
- File system as writable memory (plans, scratchpads, intermediate artifacts).
- Tool-mediated retrieval: "not only through just simple search like RAG, but all the way through tool calling."
- Iterative planning without full-context bloat.
What's missing to generalise¶
- A tool runtime with real auth (see tool-runtime-vs-harness).
- Delegated scopes per action (delegated-agent-authorization).
- Multi-user identity.
Why this matters¶
Chase/Partee are the "stop inventing new architectures" voice in 2026. The blocker isn't agent intelligence; it's the boring integration layer — OAuth, scopes, third-party service APIs. Everything else is already solved in Claude Code, Deep Agents, and their peers.
Connects to¶
- four-collaboration-zones — Sanchez's structural model for the same collaboration surface.