Create chat and worker agents with the model, instructions, tools, knowledge, and policy that define how they operate. Publish only after the required governance checks pass.
The harness brings the work of designing and governing an agent into one operating surface. Configure the model, instructions, tools, memory, knowledge sources, and policies it needs to operate, then assign owners, assess its impact, and set the approval gates required to publish. Each release retains the decisions behind it, configuration changes, gate evaluations, run history, and lifecycle records, so engineering and compliance stay aligned from first draft through production. There is no separate governance handoff after the build; governance evolves with the agent.
Build interactive chat agents with your company or product knowledge, or worker agents for jobs triggered manually, in Slack, on a schedule, by webhook, API, workflow step, or file upload.
Ground an agent in approved documents, choose its tools, and restrict it to knowledge data when the task demands a bounded answer.
Every agent runs under a policy template that combines data, access, spend, and violation rules. It can lock models, limit tools and sources, require human approval, and determine how a violation is handled.
Draft agents do not become Live until required pre-flight checks pass. Published agents retain run history, gate evaluations, and lifecycle changes.
Per-policy model locks, daily spend caps, and alert thresholds set agent guardrails. Alerts warn as a threshold is reached; a spend cap blocks the run.
Daily and monthly budgets notify the people responsible. A monthly hard limit blocks new worker runs until the limit is raised or the period resets.
Each trace records run status, captured inputs and outputs, context retrieval, model and tool activity, token use, estimated cost, duration, and guardrail outcomes. Evidence reports can be generated from the completed trace when a result needs review.
Start with an agent mode and business purpose, then define the responsible team, policy, trigger, model, instructions, tools, data sources, and operational limits it needs to work safely.
Test in the agent canvas with files and real context, then put the agent to work when appropriate. Runs can start manually, from Slack, on a schedule, by webhook, API, or workflow.
Inspect the run history and governance record beside the agent. Review inputs, context retrieval, model and tool activity, guardrail outcomes, risks, evaluations, data sources, and gate history in one place.