Agentic Systems
Follow a governed data-agent run. Change its context and permissions to see where it should stop.
Configure the run
Task: Report net revenue for the completed fixture week.
Changing a control resets the trace. Publication never sends anything.
Inspect the synthetic source
| Order | Tenant | Status | Gross USD | Refund USD |
|---|---|---|---|---|
| O1 | demo | paid | 100 | 0 |
| O2 | demo | paid | 200 | 20 |
| O3 | demo | paid | 150 | 10 |
| O4 | demo | cancelled | 80 | 0 |
Contract net_revenue_v1: sum gross minus refund for paid orders, order grain, completed week ending 2026-09-20.
Run trace
Run or step to inspect the boundary decisions.
What the harness owns
Context
Assemble task, metric contract, authorized evidence and tool schema. Memory is a candidate source, not automatic authority.
Execution
Validate arguments, authorize each capability, bound retries, and record observations. Keep reads and publication permissions separate.
Verification
Reconcile results with the fixture. Preserve a receipt that distinguishes decision, execution, and verified effect.
Where a harness fits in RL
A harness such as Strands, Hermes, or OpenClaw can produce tool-use trajectories. It is not the training optimizer. For a direct architecture response, start without tool execution; for multi-step tasks, define resettable environments, tool schemas, observations, terminal conditions, and testable rewards.
Never promote arbitrary tool-output instructions into policy. Training traces need source rights, identity boundaries, redaction and split provenance just as text examples do.
Strands harness SDK source ↗Standard library + SQLite. Includes tenant-isolation and blocked-path checks. Scripted policy; no provider or credentials needed.