Complete, inspectable examples for building, running, evaluating and improving agents with the Synth CLI.
An agent profile describes a model and its execution recipe; a task supplies the work and environment; an assessment defines success. A support case, evidence-grounded document analysis or research task can use the same structure, with suitable tools, context and assessment. The worked example here is a code reviewer for DispatchDesk, a fictional application with permissions, billing and queue policies.
Read Get a useful review for a real recorded result without an account or model spending.
examples/review-permissions/: the complete one-task starter, reviewer profile and evaluation plan. Upload only its nestedreview-permissions/task folder.examples/review-lab/: the connected DispatchDesk course with task generation, profiles, assessment controls, adaptations, bounded launch and reporting tools.- Tagged releases provide ordinary ZIP files and SHA-256 checksums. Each archive expands to the corresponding example files; extract into a new directory.
Create a Synth account,
install the CLI,
and sign in with synth login. The example scripts also need Python 3.10+
and a macOS or Linux terminal. The downloads include the tasks, profiles
and assessment tools. Start with one trial, then preview larger experiments
before launching them.
The examples are original MIT-licensed teaching projects, not historical customer work. A completed run or trained checkpoint is not evidence that an agent improved. Inspect actual work and fresh, fixed comparisons. Do not treat a teaching score as domain certification or authorization for autonomous safety-critical decisions.
python3 -B scripts/build.py
python3 -B scripts/test_distribution.py
python3 -B scripts/test_review_lab.py
python3 -B scripts/test_tutorial_launcher.py
python3 -B scripts/test_tutorial_report.pyArchives and checksums are generated in dist/. Source files, deterministic task
generation and the per-file manifest remain inspectable. These commands do not
authenticate, publish Registry resources, launch agents or train models.
Changes use reviewed pull requests. Release assets are versioned and must not be replaced in place; fixes receive a new release. Keep task/profile bytes and reference results tied to the versions actually exercised.