Agent Workflows

Define a goal, scope the agent’s tools, and set its policies. The workbench orchestrates the run and makes it repeatable.

Projects, runs, and versioned experiments

• Projects are the top-level container for code, prompts, and agent configurations.
• Runs capture a single execution, including inputs, outputs, and logs.
• Background agents keep long tasks moving while you focus elsewhere.
• Versioned experiments let you test multiple approaches in parallel.

What every workflow needs: outcome, scoped tools, limits, review

• Clear outcome (merge a PR, generate a report, update a spec).
• Scoped tools so the agent can act but not overreach.
• Profile and limits to control cost and risk in the runtime sidecar.
• Review point before changes reach production systems.

Parallel Agents

Traditional AI interfaces present single-threaded conversation. Real projects involve exploration, comparison, and parallel investigation.

• Spawning sub-agents: A primary agent working on a task can spawn sub-agents to research alternatives, investigate dependencies, or draft tests. Each runs independently and reports results back.
• Forking for exploration: Facing a decision, fork the context and direct each fork down a different path. Both develop in parallel. Compare results and pick the winner.
• Spatial overview: The workbench presents parallel activity visually, showing all active agents, their status, recent outputs, and relationships.

Reproducible behavior when agents touch production

Teams need predictability when agents touch real systems. You can reproduce any run, compare it against alternatives, and review changes before they reach production.