Agent Harness Wiki

4 min read

Prompt Optimizer Regimes for Harnesses

Goal

Turn the newly covered systems into a practical regime map for agent harnesses: when is the right abstraction runtime prompt editing, RL over a modular prompt program, in-context black-box search, evolutionary prompt search, or planning over prompt states?

Short answer

The systems split cleanly by what they treat as the writable artifact and by how much structure they assume in the feedback signal.

  • tempera — runtime editing of the live prompt instance
  • autodspy — RL over modular DSPy program structure
  • opro — black-box optimization over scored candidate histories
  • promptbreeder — evolutionary search over prompts and mutators
  • promptagent — planning over prompt states with reflective error feedback

That is not merely a taxonomy. It is a design choice about what the harness can observe, score, retain, and safely promote.

Regime map

1. Runtime adaptation: tempera

Use this regime when the problem is query-local and the harness wants to patch the active prompt without pretending it learned a durable artifact. TEMPERA is the clean example because it edits interpretable prompt components at test time: instruction phrases, exemplars, and verbalizers.

Best fit:

  • narrow tasks
  • fast local adaptation
  • human-seeded prompt skeletons
  • low appetite for persistent self-modification

Main limit:

  • it improves the live prompt instance, not necessarily the harness’s long-term artifact library

2. RL over modular prompt programs: autodspy

Use this regime when the writable object is already a structured LM workflow rather than one prompt string. AutoDSPy matters because it treats module choice, signatures, and execution strategy as the RL target.

Best fit:

  • modular LM programs
  • reusable workflow artifacts
  • evaluator loops that can score whole-pipeline behavior
  • systems that want promotion and rollback at the program level

Main limit:

  • still an early line with thinner empirical depth than the simpler prompt-search papers

3. Black-box candidate search: opro

Use this regime when the harness can score candidates reliably but does not have much structured critique. OPRO is the clean baseline: candidate/value history in, next candidate out.

Best fit:

  • scoreable artifacts
  • cheap evaluation loops
  • minimal structured trace information
  • simple optimizer baselines

Main limit:

  • weak local credit assignment; it knows what scored well, not necessarily why

4. Evolutionary search over prompts and mutators: promptbreeder

Use this regime when diversity matters and when the optimizer itself should become a writable artifact. Promptbreeder’s key move is that the mutation prompts are optimized too.

Best fit:

  • exploration-heavy search
  • robustness through diversity
  • systems willing to retain mutators, not only task prompts
  • long-lived libraries of optimization heuristics

Main limit:

  • search can become expensive and messy without a disciplined promotion policy

5. Planning over prompt states: promptagent

Use this regime when the harness already produces richer failure evidence and can afford guided search rather than flat proposal/scoring. PromptAgent is the clearest example because it treats prompt improvement as strategic planning over prompt states.

Best fit:

  • rich error traces
  • expensive but high-value prompt edits
  • domains where reflection materially improves the next move
  • systems that already preserve intermediate failure evidence

Main limit:

  • more machinery, more search budget, and more dependence on the quality of reflective feedback

My synthesis

The real divider is not RL versus non-RL

The sharper divider is what kind of learning signal the harness has.

Durable harnesses should store more than a “best prompt”

These systems collectively imply that a serious harness may want to retain:

  • candidate prompt histories
  • reflective error analyses
  • mutation prompts and search heuristics
  • module-selection policies
  • promotion / rollback provenance

That begins to look less like prompt engineering and more like release engineering for language programs.

The compile-time versus run-time boundary still matters

tempera is the clearest runtime-adaptation case here. autodspy, opro, promptbreeder, and promptagent all lean more toward artifact search that could, in principle, feed a durable release pipeline. A harness that blurs these lanes risks mistaking attractive improvisation for real learning.

Bottom line

The new coverage fills three missing slots in the regime map:

  1. runtime prompt adaptation via tempera
  2. RL over modular prompt programs via autodspy
  3. three non-identical search baselines via opro, promptbreeder, and promptagent

Together with rlprompt, dspy, dspy-assertions, sammo, textgrad, and gepa, the wiki now has a much cleaner optimizer-family map for prompt programs and language workflows.

Read this with prompt-optimization-and-dspy-follow-ups, prompt-program-representation-and-optimizer-open-questions, prompt-program-deployment-open-questions, and prompt-optimization-timeline-and-harness-lessons.

Graph View

Backlinks