Agent Harness Engineering — Stack · Agent Architecture and Orchestration

Legacy handbook baseline: May 2026. Source-verified additions reviewed 2026-10-13. The sources below establish these implementation examples, not a complete market ranking.

Selection criteria

Inspect the mechanisms behind a runtime's convenience interface: control over context projection, durable work records, tool dispatch permissions, workspace isolation, and checker placement. Ask which configuration is captured with a run and whether it can resume after the application process is replaced. Probe the recovery boundary before selecting on demos.

Implementation examples

Option	Useful starting point	What remains application work
Deep Agents	A packaged harness for planning, filesystem work, and delegated tasks	Define the allowed operations and what proves your objective complete
Anthropic's long-running-agent example	Environment initialization, persistent progress, and incremental verification	Adapt the artifact and checks to your workload; a demonstrated coding setup is not a universal recipe
A small application-owned runtime	Few tools and precise control over each boundary	Implement context handling, durable records, and recovery explicitly

Framework adoption does not transfer responsibility for credential scope or outcome verification. Treat defaults as configuration to inspect. Keep runtime, model, prompt, tool-contract, and checker revisions in the experiment record so results remain attributable.