AgentBeat’s primary integration is a small HTTP contract. SDKs are optional helpers for building that Target boundary and collecting L2 Evidence; they are not alternate evaluation engines. For framework-specific wiring and the newly packaged adapter kits, continue to SDK integration.

​
Default Connector

The Go Core registers business-http-json-v1 as the declarative HTTP Connector.
RouteRequiredResponsibility
GET /healthzDeployment health checksReport Target readiness
POST /v1/agent/invocationsL1/L2/L3Accept one target-invocation-v1 task and return the final response
GET /v1/evidence/{run_id}Optional L2/L3Return same-run target-evidence-v1 messages and tools
The Connector rejects redirects, enforces response-size limits, checks schema/run/case binding, and requires a non-empty final output. Endpoint and headers come from the selected Registry Deployment. See the complete L1 request and Registry example →

​
No SDK is required for L1

Any service can implement the HTTP contract directly. The Agent only receives the trusted task text and evaluator-managed correlation metadata; the Case’s attack fields, expected result, Judge rubric, endpoints, and credentials are not sent in the business request. 
AgentBeat Go Core                 Customer Target
──────────────────                ───────────────
registered endpoint   ── POST ──▶ /v1/agent/invocations
run_id + case_id                  existing Agent logic
trusted task          ◀─ JSON ── final_response

​
JavaScript SDK

OPTIONAL INTEGRATION agentbeat-sdk-js provides a Target server and bounded Evidence collector: 
import { createTargetServer } from "agentbeat-sdk/node";

const server = createTargetServer({
  targetId: "target-my-agent",
  auth: targetBearerToken,
  evidence: { source: "my-agent", observedChannels: ["messages", "tools"] },
  async invoke({ input, observe }) {
    const result = await myExistingAgent(input.text);
    observe?.message({ role: "assistant", text: result.answer });
    return { finalResponse: result.answer };
  },
});

server.listen(8091, "0.0.0.0");
Package responsibilities:
  • agentbeat-sdk: EvidenceCollector, bounded storage, Evidence building;
  • agentbeat-sdk/node: Target HTTP server and bearer-token helper;
  • agentbeat-sdk/adapters/codex-app-server: notification-to-Evidence projection;
  • agentbeat-sdk/legacy: compatibility only; do not use for new integrations.

​
Python SDK

OPTIONAL INTEGRATION agentbeat-sdk-py exposes the same narrow Target-side role: 
from agentbeat_sdk import create_target_server


def invoke(context):
    answer = my_existing_agent(context.text)
    if context.observe is not None:
        context.observe.message(role="assistant", text=answer)
    return {"final_response": answer}


server = create_target_server(
    target_id="target-my-agent",
    auth=target_bearer_token,
    invoke=invoke,
    evidence={"source": "my-agent", "observed_channels": ["messages", "tools"]},
)
server.serve_forever()
The Python package also includes a LangGraphObserver that projects completed stream updates into the optional collector. It does not replace LangGraph models, tools, checkpoints, or callbacks.

​
L2 Evidence boundary

The optional collectors produce bounded, redacted target-evidence-v1 documents. They can observe messages and paired tool call/results; an optional files channel can carry validated file-change events at the Target protocol boundary. The Go evaluator requires the Target Evidence document to declare messages and tools, validates event order and bindings, and projects the Evidence channels required by its L2 policy. A missing required channel remains unavailable.

​
L3 is not an Agent adapter

A State Controller belongs to the environment side and is registered separately. It owns reset, before/after snapshot, and native verification. Do not put Controller authority inside the Agent SDK.

​
Legacy compatibility

LEGACY COMPATIBILITY
  • sdk/agentbeat-cli keeps an older SDK-hosted adapter flow working.
  • core/go/cmd/agentbeat is a compatibility adapter.
  • New evaluation integrations use the Go Core through agentbeat eval-run.
  • Inspect-based benchmark work is not the primary AgentBeat Target runtime.

Observation model

Evidence and scoring