AgentClash

Synthetic data

Synthetic data generation for AI agents that teaches, not noise

Bulk self-instruct produces volume without signal. AgentClash and DataSmith run a weak-vs-strong judge loop that accepts examples only when a strong solver succeeds and a weak solver struggles: the useful difficulty zone for fine-tuning and eval.

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

Why weak-vs-strong beats prompt-only generation

Built for reviewable agent decisions

Meta FAIR Autodata showed that agentic self-instruct with solver separation beats standard synthetic pipelines across CS, legal, and math settings. DataSmith implements the practical loop; AgentClash runs it hosted for regression-ready datasets.

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

From seeds to export or CI gates

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

Synthetic data generation FAQ

What is synthetic data generation for AI agents?

It is the process of creating labeled training or eval examples for agent workloads using LLM-driven pipelines instead of manual curation alone.

How does Agentic Self-Instruct improve data quality?

A challenger proposes examples, weak and strong solvers attempt them, and a judge accepts rows only when the score gap indicates the example is learnable but not trivial.

Should I use DataSmith or AgentClash for generation?

Use DataSmith for offline SFT, DPO, and Hugging Face export. Use AgentClash when generated examples should feed dataset evals, baselines, and CI regression gates.