AgentClash

Agentic Self-Instruct

Agentic Self-Instruct: weak-vs-strong synthetic data

Agentic Self-Instruct is a four-role loop for synthetic dataset generation: challenger proposes, weak and strong solvers attempt, judge filters. DataSmith ships it as a Python SDK; AgentClash runs it hosted for eval-ready datasets.

live eval
gate: pass

Candidate

92correct patch, low cost

Baseline

88stable reference run

Control

73missed edge case

replay timeline

1loaded task inputs and tool policy
2ran sandbox actions and captured artifacts
3scored trajectory and validator evidence
4attached scorecard and release verdict

ci verdict

Candidate clears release gate

Correctness improved, latency within budget, and required artifacts were preserved for review.

agentclash run create --follow

The four roles

Built for reviewable agent decisions

Challenger generates candidates from seeds. Weak solver represents the model you want to improve. Strong solver represents a reference path. Judge scores quality and separation before acceptance.

OpenTelemetry-compatible trace import

Pinned datasets and golden test cases

Baseline versus candidate regression checks

Replay trails for tool calls, outputs, and artifacts

Scorecards for correctness, cost, latency, and evidence

CI gates for prompt, model, RAG, and tool changes

Workflow

Research to production

Import the evidence

Start from OpenTelemetry traces, curated datasets, support transcripts, or a real failure your team already saw.

Pin the baseline

Record the current accepted behavior so every prompt, model, RAG, or tool change has a fair comparison point.

Replay the evidence

Inspect tool calls, outputs, artifacts, latency, cost, and judge evidence when a candidate gets worse.

Gate the release

Compare candidate and baseline runs, then fail CI before a regression reaches users.

FAQ

Agentic Self-Instruct FAQ

What is Agentic Self-Instruct?

An agentic loop where a challenger generates examples and weak/strong solvers plus a judge determine whether each row is in the useful difficulty zone for training.

How is it related to Meta FAIR Autodata?

Autodata formalized weak-vs-strong agentic self-instruct for synthetic data. DataSmith implements the practical inner loop; it is not an official Meta release.

Can I tune acceptance thresholds?

Yes. Both DataSmith and AgentClash expose strong pass rate, weak fail rate, and minimum score gap controls.