Pradra StudioDocs
Sign in

Agents

An agent is a capability that pursues a goal instead of answering a single question: given "resolve this billing dispute", it plans, calls tools, observes the results, and works toward an answer — deciding its own steps as it goes. Pradra Studio runs agents inside a strict harness: budgeted, fully traced, and gated like every other capability.

Guardrails, checked every turn

An agent's configuration sets hard limits, and every one is enforced on every step — in a fixed order, with no exceptions:

  • Wall-clock time and step budget — a loop that isn't converging is stopped, with its full trace intact.
  • Cost per run and per day, on top of the shared monthly budget.
  • Tool allowlist — the agent can only call tools its capability was explicitly granted. An ungranted tool is invisible to it.
  • Approval for consequential actions — a tool call that changes the outside world always parks for a human; see Agent tools & approvals.

A refusal (blocked tool, exhausted budget) is fed back to the agent as an observation — it can adapt its plan — but the limit itself never moves.

No trace, no run

Every step — each plan, tool call, observation, and violation — is persisted before the run proceeds. There is no such thing as an agent run you can't reconstruct afterward. The trace viewer shows the whole conversation between the agent and its tools, step by step.

Two more honesty rules:

  • Numbers the agent's final answer reports are copied verbatim from tool observations where possible — the model may summarize, but measured values are pinned, not paraphrased.
  • A final answer that doesn't match the capability's output contract is a failed step, not a delivered answer — the agent keeps trying within its budget.

Evaluated before it can serve

Agents pass the same kind of gate as everything else: a golden set of goals is run through the real loop (with side effects disabled), scoring task success, steps taken, cost, and violations. An agent with no passing evaluation — or one that fails your acceptance criteria — cannot be activated. In production, agent health reports insufficient evidence under low traffic rather than a fabricated "healthy".

Working with agents in the Studio

  1. Create a capability with type agent inside a solution; its detail page gains tabs: Model (the LLM behind it), Guardrails, Tools (grants), Runs, and Evaluation.
  2. Grant it tools from the tool registry.
  3. Test with runs; open any run's trace to watch its reasoning.
  4. Build the golden goal set, evaluate, and activate through the gate.

Next