← ClaudeAtlas

whileai-simulationslisted

Use the While simulations SDK to generate agent trajectories from a policy, tools, an agent harness, seed tasks, or production traces; focus generation on failures, grade with the developer's own authority, inspect quality, and export usable datasets. Use when a coding agent is asked to simulate an agent, expand coverage, turn traces into new scenarios, create eval or post-training data, or run the simulate-grade-select loop.
whilehq/whileai-sdk · ★ 3 · AI & Automation · score 77
Install: claude install-skill whilehq/whileai-sdk
# While simulations Operate While for the developer. Discover the context already present in their repository, choose the shortest valid input path, run a small smoke test, and return auditable data plus a run report. Do not require every possible input: policy, tools, traces, seeds, and a harness are complementary. Terms in this skill: the **developer** is the person or team using While; the **coding agent** is the assistant reading this skill and operating the SDK; the **target agent** is the agent being simulated and improved; the **simulated user** is the person speaking to the target agent inside a trajectory. Ground truth is the installed `whileai.simulations` package. Inspect its public signatures when the installed version differs from these examples. ## Inputs and routes Use what the developer has: - `tools=` plus `system_prompt=`: cold-start simulation across the declared agent and policy space. - `traces=` plus the agent definition: mine observed failures and concentrate search on their tools, faults, world states, and behaviors. About 20 graded or fault-bearing traces gives useful targeting; below 10, treat the result mostly as cold-start exploration and say so. Traces reproduce situations (tools, faults, world states); a failure that lives in how the reply is worded has no world-visible trigger, so pass `grader=` as well and the loop mutates on graded failures too (`search["mutation_aims"]`). - `agent=`: connect an existing callable, supported