C-03 Agent Loop Controller

Brick C: Agents, autonomy, execution. French version: C-03-agent-loop-controller

1. Role and responsibilities

Drives an agent turn with enterprise_ai.runtime.AgentRunner: model call, then tool calls, until the final answer, under budgets (steps, tool calls, cost, duration). The Navin continuation rules (enterprise_ai.runtime.continuation.default_rules) prompt the model again when work remains (remaining_work), when the user already said go (go_ahead), when it must verify before finishing (verify_before_done, verify_failed), when the answer is empty or truncated (empty_final_response, length_recovery) or makes no progress (no_progress, wrap_up). The run plan is carried by enterprise_ai.loop.plan.RunPlan, which starts the next ready step automatically.

2. Position in the architecture and interactions

• Called by AgentLoop.run_once() for each message or pending run.
• Checks the heartbeat (POST /v1/runs/{id}/heartbeat) before every model or tool call and stops on action=stop.
• Records steps (POST /v1/runs/{id}/steps) and checkpoints.

3. Open source component and developed part

• Engine: Navin architecture of the Python SDK, with no declared graph. enterprise_ai.loop.AgentLoop takes in messages, prepares the context, manages sessions, resumes a run after a crash, handles stop, cron and heartbeat, and exposes serve(); enterprise_ai.runtime.AgentRunner runs one turn; enterprise_ai.subagents.SubagentManager runs sub-agents in parallel; enterprise_ai.loop.plan.RunPlan carries the plan; enterprise_ai.runtime.bus.MessageBus links channels and engine.
• Optional adapters: LangGraph (enterprise_ai.runtime.langgraph.build_graph) and Temporal (enterprise_ai.runtime.temporal.build_workflow) reuse the same steps; neither is required.
• Server side, python -m eai.worker runs the periodic jobs: scheduler tick, event outbox drain, quota reservation expiry, stale heartbeat detection, retention.

4. Data model

Tables (schema eai, tenant_id RLS):

• eai.run_steps
• eai.runs

Common columns: id uuid, tenant_id uuid NOT NULL, created_at, updated_at; RLS policy current_setting('app.tenant_id', true)::uuid.

5. Contract

• Applied contract: execution-contract
• JSON Schema: contracts/v1/execution-contract.schema.json
• Key fields: steps[].kind, steps[].status, budgets.max_steps, budgets.used
• Validation: POST /v1/contracts/{name}/validate; manifest apply: POST /v1/manifests/apply.

6. API endpoints

• POST /v1/runs/{id}/steps
• GET /v1/runs/{id}/steps
• POST /v1/context/plan

Request POST /v1/runs/{id}/steps:

{
  "kind": "tool_call",
  "name": "jira.search",
  "input": {
    "q": "INC-42"
  },
  "fingerprint": "sha256:9ab..."
}

Response:

{
  "step_index": 4,
  "status": "recorded",
  "budgets": {
    "steps_used": 4,
    "max_steps": 20
  }
}

7. SDK examples

Python:

from enterprise_ai import EnterpriseAI
from enterprise_ai.loop import AgentLoop
from enterprise_ai.loop.plan import RunPlan

client = EnterpriseAI(base_url="<control-plane-url>", api_key="eai_xxx_secret")
plan = RunPlan()
plan.add("collect", "Collect the ticket history")
plan.add("summarize", "Write the summary", depends_on=["collect"])
loop = AgentLoop(client=client, model="claude-sonnet-eu", agent_id="agt_support")
result = loop.run_once("Summarize INC-42")
plan.complete("collect", evidence="12 messages read")   # the next ready step starts on its own

TypeScript:

import { EnterpriseAI, AgentRunner } from "@enterprise-ai/sdk";

const client = new EnterpriseAI({ baseUrl: "<control-plane-url>", apiKey: process.env.EAI_API_KEY });
const runner = new AgentRunner({ client, model: "claude-sonnet-eu", agentId: "agt_support" });
const turn = await runner.runTurn([{ role: "user", content: "Summarize INC-42" }]);

Java:

EnterpriseAIClient client = EnterpriseAIClient.builder()
    .baseUrl("<control-plane-url>")
    .apiKey(System.getenv("EAI_API_KEY"))
    .build();
var run = client.runs().start("agt_support", Map.of("task", "Summarize INC-42"));
var answer = client.heartbeats().beat(run.get("id").asText(), Map.of("step", "plan"));

Go:

c, _ := eai.NewClient(eai.WithBaseURL("<control-plane-url>"), eai.WithAPIKey(os.Getenv("EAI_API_KEY")))
run, _ := c.Runs.Start(ctx, "agt_support", map[string]any{"task": "Summarize INC-42"})
answer, _ := c.Heartbeats.Heartbeat(ctx, "run_01J9", map[string]any{"step": "plan"})

8. Configuration

• Environment variables: EAI_DATABASE_URL
• Helm keys: controlPlane.config
• Client side: AgentRunner and LoopBudget budgets (steps, tool calls, cost, duration)

9. Security controls

• Every tool action goes through the Tool Gateway (policy, DLP)
• spec.autonomy level bounds actions without approval
• Every request carries the tenant resolved from the credential; PostgreSQL RLS
• Every mutation and policy decision writes a hash-chained audit event

10. Observability

• Prometheus metrics: eai_loop_steps_total, eai_loop_detected_total, eai_loop_stop_reason_total{reason}
• Traces: OpenTelemetry spans with GenAI semantic conventions (gen_ai.operation.name, gen_ai.agent.id, gen_ai.tool.name), attributes eai.tenant_id, eai.run_id, traceparent propagation.
• Audit actions: run.transition

11. FinOps units

• Tokens per step, cumulative cost per run.
• Emitted as Usage Event (POST /v1/finops/usage) with tenant, project, agent and run.

12. Failure modes and resilience

• Loop without progress: the no_progress then wrap_up rules force a conclusion
• Budget reached: the turn stops with the reason of the limit reached
• Stop requested by heartbeat or kill switch: stop before the next call

13. Standalone adoption (client of a client)

• enterprise_ai.agent (local loop) + model gateway only.
• SDK: enterprise_ai.agent (AgentsClient, RunsClient, SessionsClient)
• Minimal deployment: Helm chart with controlPlane.enabled=true and PostgreSQL; disabled external components use their in-process fallback.

14. Acceptance tests

1. 3 identical calls -> loop_detected stop
2. max_steps reached -> state failed budget_exceeded
3. Cross-tenant access refused
4. Error returned in fr with Accept-Language: fr