The engine behind
every intent call.

You call an intent. The orchestrator decides whether to use single-shot or multi-step agentic execution, which tools to invoke, and how many reasoning steps to take.

You never manage prompts, tool schemas, or agent loops. You get structured output and a full execution trace in every response.

# POST /v1/run
{
  "pack": "talent",
  "intent": "analyze-resume",
  "input": { "resume_text": "..." }
}

# Response
{
  "response": {
    "content": "Candidate scores 87/100...",
    "intent": "analyze-resume",
    "confidence": 0.93,
    "metadata": { "guardrails_passed": true }
  },
  "execution": {
    "steps": 3,
    "latency_ms": 390,
    "agent_mode": "agentic",
    "tool_calls": [
      "fetch_profile",
      "score_criteria",
      "rank_fit"
    ]
  }
}

Automatic mode selection

The orchestrator inspects intent complexity and routes to single-shot LLM or multi-step agentic execution — no configuration needed. You call the intent; we pick the right strategy.

Managed tool registry

Domain Packs ship with pre-registered tools scoped to their vertical — DB lookups, document retrieval, scoring functions, and external APIs. The engine calls them; you see the results in the trace.

Built-in error recovery

Retry logic, exponential backoff, fallback paths, and partial-result surfacing are handled at the engine layer. Agents recover silently; you never write recovery code.

Session memory & context

The orchestrator maintains a rolling context window per session — prior turns, retrieved documents, and intermediate reasoning — injected automatically into each agent step.

Guardrails at every step

Policy checks run before and after each tool call, not just on final output. The engine enforces your domain's guardrail policies across the entire execution graph.

Full execution trace

Every intent call returns a structured trace: steps taken, tool calls made, latency per step, agent mode used, and guardrail results. Query it in the dashboard or stream it to your observability stack.