01
Standard agents
Step-by-step with explicit human approval at each significant action. Best for high-stakes workflows where every move should be sanctioned.
Agents execute multi-step tasks, fetching data, calling tools, making decisions, and pause for human approval whenever risk crosses your threshold. Built on a ReAct loop with end-to-end observability, every step logged.
Not every task wants the same leash. Configure each agent for the operating envelope that fits the use case it's bound to.
Step-by-step with explicit human approval at each significant action. Best for high-stakes workflows where every move should be sanctioned.
Run to completion within the policy envelope. Hit defined approval gates only when crossing risk thresholds. Best for high-volume, well-bounded work.
Orchestrate other agents from inside Symphony workflows. Plan the sequence, dispatch sub-agents, gather their outputs, and decide what's next.
Email send, document fetch, database query, web search, calendar lookup, HRIS read, framework reference. Tools are policy-checked at invocation time.
Per-step heartbeat with token usage, tool calls, LLM reasoning preview, and duration. See what your agent is doing, in real time.
One click to stop a run mid-flight, sanction a pending action, or block one. Every decision lands in the audit log automatically.
Every agent is bound to a use case in Governance 1st. That's where its policies, approvals, evaluators, and risk tier live, so an agent can't quietly wander outside what was sanctioned.
An agent without a registered use case can't be deployed. No shadow agents, ever.
The use case's risk score drives the agent's approval thresholds and evaluator coverage.
Agents only see the tools the use case sanctioned. Adding a tool means re-approval.
Every step, tool call, and approval decision lives in the same log auditors review.
Ten behavior checks per LLM call: hallucination, drift, bias, dangerous advice, injection, refusal calibration, bigotry, logical fallacy, partisanship, misinformation.
Disable a use case and every agent bound to it stops cold. Same pattern across the platform.
30-minute working session: bring a real workflow, we configure an agent against it, register the use case, and run it with full observability and approval gates.