Automated Agent Optimization Taxonomy

Classify automated agent-improvement methods by what evidence the optimizer observes and what surface it is allowed to edit, to see where full-harness evolution sits relative to single-surface methods.

Automated methods for improving an LLM agent without retraining the base model form a spectrum, organized by two axes: what evidence the optimizer observes, and what surface it is allowed to edit. Placing a method on this map clarifies what kind of gain to expect from it and what it will miss:

Episodic critique / reflection — observes the agent's own recent output or trajectory; edits nothing durable, only the next attempt's in-context state. Reflexion and Self-Refine-style self-critique prompting live here: fast, cheap, but the lesson does not automatically persist past the episode unless paired with episodic reflection memory.
Prompt / instruction optimization — observes a batch of trajectories or labeled examples across many episodes; edits a single prompt-level surface (instructions, a natural-language playbook, or few-shot demonstrations). TextGrad-style prompt optimization tooling and DSPy teleprompters live here, along with reflective playbook distillation and trajectory-feedback RL variants that reinforce successful tool sequences without touching tools or middleware themselves.
Program-structure editing — observes trajectories the same way, but edits the shape of the system: a growing skill library, a scored archive of program variants under mutation, or a searched workflow graph. This is where a multi-component harness-evolution loop belongs — it jointly edits system prompt, tools, middleware, and long-term memory as a combinatorial whole, rather than one editable surface, so cross-component trade-offs become legible to the optimizer instead of hidden inside whichever surface happens to be editable.

The practical payoff of placing a method correctly: a method confined to the prompt-optimization tier cannot fix a failure that only a tool-implementation or middleware-level guard can close (see harness component-level selection) — no amount of instruction rewriting substitutes for an execution-time intercept. Conversely, a full harness-evolution loop pays for that reach with a larger, more heterogeneous action space and a harder attribution problem (see evidence-driven change manifest), which is exactly the coordination cost prompt-only methods avoid by staying on one surface.

Sources

Agentic Harness Engineering (Lin, Liu, Pan, et al.), §2