Skip to main content
An agent can do well on a public benchmark and still miss the things your team cares about. A support agent must follow your escalation policy; a document analyst must ground its conclusions in the supplied evidence; an engineering assistant must apply the right project requirements. Synth connects that work to an improvement loop: prepare representative tasks, define how to assess the results, evaluate an agent, train its model, and test the resulting checkpoint on different work. You keep the configurations and evidence behind each decision. An agent profile defines the model and execution recipe; a task supplies the work and environment; a verifier or judge assesses the result. Use these ingredients for coding, research, document processing, customer support or another workflow, with tools and assessment suited to your work.

Try a complete example

Follow a code-review example from source to finding, then download the inputs and run it yourself.

​
Start with your goal

Run an agent

Execute a prepared task, collect the output, and understand the result.

Measure an agent

Compare setups on the same work. Inspect quality, reliability and failures.

Improve an agent

Train on relevant tasks, then compare the saved checkpoint with the baseline on held-out work.
You do not need to start with training. A baseline can reveal that the real problem is missing context, an unreliable tool or a scoring rule that rewards the wrong behavior.

​
Bring work that matters

Good tasks resemble situations in which you want to trust the agent: review a change, resolve a support case, reconcile a report against source records, or draft a project-specific checklist. They include the information and tools the agent would normally have, plus a way to assess the result. For safety-critical work, such as civil engineering, specialist validation and appropriate human approval remain necessary; an evaluation score is not professional certification. A task is reusable. Its exact version stays fixed while you compare agents or repeat an experiment. A dataset keeps a selection of those task versions together so the comparison does not quietly change underneath you. Start with what a realistic task needs and how to assess its result.

​
Keep the agent and the model distinct

An agent combines a model, a harness and the context available during its work. A profile records the model and execution recipe: instructions, files, setup, tools and execution behavior. A familiar tool such as Claude Code is a harness, not a promise about which model runs inside it. Training a supported open model through that harness does not train Anthropic’s model. Understand agent profiles before comparing or changing them.

​
Read the evidence

The useful output is more than a score. Inspect the review, patch or other work the agent produced, how it reached that result, and why the assessment passed or failed. A run can finish successfully while the agent performs poorly. Likewise, saving a trained checkpoint does not establish that the agent improved. Read the results and compare on held-out work before deciding what to use.

​
Fit Synth into your workflow

Native agent recipes

Keep an explicit, versioned execution recipe instead of reconstructing an experiment from terminal history.

Harbor task packages

Reuse a compatible task format for instructions, environments and verification.

CLI and automation

Use the same commands interactively and in scripts. Read structured JSON; no Synth SDK is required.
New here? Explore the complete example, then install and sign in when you are ready to run it. For the overall model, read How Synth fits together. If you already have a task and profile, use the existing-input quickstart.