Effective agent capability is dominated by environment design (tools, prompts, state injection, deterministic pre-compute), not by model selection — in the regime where the candidate model is capability-sufficient for individual turns.

Evidence: Anthropic's 'Building Effective Agents' (tool design as the dominant lever); empirical findings from a 600-LOC ReAct harness against a 9B model where state-shaped progress blocks dominated every other intervention; a prospector build where cross-parcel bundling via derived columns produced a multiplicative quality lift; MTRouter consistent-with as a third-party data point.

Revision condition: A clean matched-environment study showing that varying model identity, holding harness fixed, produces variance comparable to or larger than varying harness, holding model fixed. That study does not yet exist in published form.

Distilled May 2026 from prior engineering work.

Default LLM agents are structurally Episodic, and operationally more Episodic than they architecturally need to be.

Evidence: Across-session amnesia is architectural; the within-session version is empirical — harnesses that inject state-shaped progress blocks substantially outperform those that rely on the agent re-deriving from execution history, even though the history is in-context and the agent could in principle use it.

Revision condition: A demonstration that, with the right prompting alone (no state injection), an agent can match the performance of a state-injection harness on a long-horizon task. I have not seen this.

Refined from the original 'AI agents lack persistent memory' framing through a series of harness experiments.