Concepts
Engines
GEPA, Best-of-N, Meta-Harness, AutoResearch, AutoSaddler, ShinkaEvolve and Auto.
An engine is the search method that proposes new versions of your artifact. Skynet has six native engines for Prompt, Code and Text artifacts, three DSPy optimizers for DSPy programs, and Auto. This page says what each one does, when to pick it, and which settings it has.
Read How optimization works first if you do not know which engines fit your artifact.
Quick comparison
| Engine | Artifact kinds | Coding agent settings | Max iterations setting |
|---|---|---|---|
| Auto | All | Yes, on Prompt, Code and Text | No |
| GEPA | Prompt, Code, Text | No | No |
| Best-of-N | Prompt, Code, Text | No | No |
| AutoResearch | Prompt, Code, Text | Yes | No |
| Meta-Harness | Prompt, Code, Text | Yes | Yes |
| AutoSaddler | Prompt, Code, Text | Yes | Yes |
| ShinkaEvolve | Prompt, Code, Text | No. It has its own editor setting. | Yes |
| DSPy GEPA | DSPy program | No | No |
| DSPy MIPROv2 | DSPy program | No | No |
| DSPy BootstrapFewShot | DSPy program | No | No |
The wizard marks GEPA, DSPy GEPA and Auto as optimizers that use the feedback text from your metric.
Auto
Auto is the default. What it runs depends on the artifact type:
- DSPy program: Auto runs DSPy GEPA.
- Prompt, Code or Text: Auto explores with five engines at the same time: GEPA, AutoResearch, Meta-Harness, AutoSaddler and ShinkaEvolve. Each engine gets an equal share of the scorer runs: the budget divided by 6. Then GEPA continues from the best version with the rest of the budget.
Auto does not use Best-of-N.
Pick it when you do not know which engine fits. This is the best place to start.
GEPA
"Reflective evolution with a Pareto front of versions." The reflection model reads the score and the feedback of each case and writes a better version. GEPA keeps a front of versions that are each best on some cases.
Pick it when your metric gives specific feedback, and you want a strong single engine.
Best-of-N
"Independent proposals from the reflection model; keep the best." Each proposal is independent of the others. Skynet re-scores the top proposals on the validation cases.
Pick it when you want a simple baseline to compare with other engines.
AutoResearch
A coding agent runs research rounds. It edits, scores and keeps what helps until the budget runs out.
Pick it when a good change needs several steps of investigation, for example on a Code artifact.
Meta-Harness
A coding agent proposes a batch of candidates each iteration. It uses the full history of candidates and scores.
Pick it when you want an agent that learns from the history of all earlier attempts. Use Candidates per iteration to set the batch size.
AutoSaddler
A coding agent reads the feedback from your metric and patches the version. Skynet keeps only patches that a new scorer run confirms.
Pick it when your metric feedback points to clear, local defects.
ShinkaEvolve
"Evolves many versions in parallel islands and learns which models improve them most." It is an evolutionary search over a population of candidates, with mutations and crossovers that a model writes.
Pick it when you want broad exploration and you have a larger budget.
DSPy optimizers
| Optimizer | What it does |
|---|---|
| DSPy GEPA | Reads the metric feedback and evolves the instructions of each predictor. |
| DSPy MIPROv2 | Proposes instructions and few-shot demos together, and searches their combinations with Bayesian optimization. |
| DSPy BootstrapFewShot | Collects worked examples that the program gets right and adds them as few-shot demos. It does not use a reflection model. |
DSPy MIPROv2 and DSPy BootstrapFewShot need pairs (cases with expected outputs).
Settings
Budget
| Setting | Default | Range |
|---|---|---|
| Max scorer runs | 200 | 1 to 100000 (Auto: at least 6) |
| Stop at score | Off | 0 or more |
| Max iterations | Off | 1 to 1000. Only Meta-Harness, AutoSaddler and ShinkaEvolve. |
One scorer run is one call of your metric on one case. The final re-score of the starting version and the winner does not count against Max scorer runs.
Coding agent settings
These settings apply to Auto, AutoResearch, Meta-Harness and AutoSaddler.
| Setting | Default | Range |
|---|---|---|
| Coding agent | Codex | Pi, Codex, OpenCode, Prime Agent, Custom command |
| Candidates per iteration | 3 | 1 to 8. Shown for Auto and Meta-Harness. |
| Max tool calls per candidate | 100 | 1 to 1000 |
ShinkaEvolve settings
| Setting | Default |
|---|---|
| How new versions are written | Single model call. The other option is "Coding agent (Pi)". |
| Islands | 2 (1 to 8) |
| Migration interval | 10 |
| Parent selection | Weighted. Other options: Power law, Beam search. |
| Diff edits / Full rewrites / Crossovers | 60% / 30% / 10%. They must add up to 100%. |
| Archive size | 40 |
| Parallel evaluations | 2 |
| Parallel proposals | 2 |
When Auto runs ShinkaEvolve, the wizard shows these settings as "ShinkaEvolve settings (when Auto runs it)".