Claude Managed Agents Integration¶
Pisama integrates with Claude Managed Agents to monitor agent sessions and detect failure modes in Anthropic-hosted agent systems.
How the integration works¶
Managed Agents runs on Anthropic's infrastructure and does not push webhooks to external systems. Pisama connects to the Anthropic API to pull session events via the /sync endpoint, then runs its detection pipeline.
Primary path (recommended): Register your Anthropic API connection, then call POST /sync to pull sessions and run detection.
Secondary path: Build your own bridge that streams from Anthropic's SSE endpoint and posts pre-processed session data to Pisama's /webhook endpoint.
Setup¶
1. Get a Bearer Token¶
curl -X POST https://api.pisama.ai/api/v1/auth/token \
-H "Content-Type: application/json" \
-d '{"api_key": "pisama_your_api_key_here"}'
2. Register a Connection¶
curl -X POST https://api.pisama.ai/api/v1/managed-agents/connections \
-H "Authorization: Bearer $TOKEN" \
-H "Content-Type: application/json" \
-d '{
"name": "production",
"api_key": "<your_anthropic_api_key>",
"organization_id": "<your_anthropic_org_id>"
}'
3. Register Agents¶
Register each agent you want to monitor. The mcp_servers and allowed_hosts fields enable platform-specific detection.
curl -X POST https://api.pisama.ai/api/v1/managed-agents/agents \
-H "Authorization: Bearer $TOKEN" \
-H "Content-Type: application/json" \
-d '{
"connection_id": "<connection_id>",
"agent_id": "<anthropic_agent_id>",
"agent_name": "My Research Agent",
"mcp_servers": ["linear", "github"],
"allowed_hosts": ["api.github.com", "*.linear.app"]
}'
4. Sync Sessions (Pull Model)¶
Pull recent sessions from the Anthropic API. This fetches session events, creates Pisama traces, and runs detection.
curl -X POST https://api.pisama.ai/api/v1/managed-agents/sync \
-H "Authorization: Bearer $TOKEN" \
-H "Content-Type: application/json" \
-d '{
"connection_id": "<connection_id>",
"limit": 20
}'
Set up a cron job to sync periodically:
*/15 * * * * curl -sS -X POST https://api.pisama.ai/api/v1/managed-agents/sync \
-H "Authorization: Bearer $TOKEN" \
-H "Content-Type: application/json" \
-d '{"connection_id": "<id>", "limit": 20}' >> /tmp/pisama-sync.log 2>&1
5. Discover Agents¶
List all agents from the Anthropic API for registration:
curl -X POST https://api.pisama.ai/api/v1/managed-agents/discover \
-H "Authorization: Bearer $TOKEN" \
-H "Content-Type: application/json" \
-d '{"connection_id": "<connection_id>"}'
Event Model¶
Pisama understands the real Managed Agents event types:
| Event Type | Category | What Pisama analyzes |
|---|---|---|
user.message | User input | Content for hallucination, injection detection |
agent.message | Agent output | Content for drift, hallucination, loop detection |
agent.thinking | Reasoning | Thinking content for coherence analysis |
agent.tool_use | Built-in tool | Tool call patterns for loop, stall detection |
agent.tool_result | Tool output | Error rates, result quality |
agent.mcp_tool_use | MCP tool | MCP server scope, retry patterns |
agent.mcp_tool_result | MCP result | MCP failure rates, error patterns |
agent.custom_tool_use | Custom tool | Custom tool call patterns |
span.model_request_end | Telemetry | Token usage, cost aggregation |
session.status_* | Lifecycle | Status transition validation |
session.error | Errors | Infrastructure error tracking |
agent.thread_message_* | Multi-agent | Coordination analysis (research preview) |
Detection Capabilities¶
Managed Agents-Specific Detectors¶
| Detector | Failure Mode | Description |
|---|---|---|
| Session Stall | F13 (Timeout) | Large event gaps, sessions stuck without progress |
| MCP Failure | F14 (Tool Failure) | MCP tool errors, retry patterns, high error rates |
| Cost Overrun | F16 (Budget) | Token aggregation from span events, budget thresholds |
| Session Corruption | F2 (State) | Invalid status transitions, timestamp regressions, session errors |
| Tool Permission | F5 (Security) | MCP tool calls to unregistered servers |
| Environment Escape | F5 (Security) | Blocked host access attempts (infrastructure prevents the action; this detects the intent) |
Core Detectors (Also Applied)¶
All 20 core detectors run on Managed Agents session content, including loop detection, hallucination, persona drift, coordination failures, and more.
OTEL Integration¶
If sessions emit OpenTelemetry spans, Pisama auto-detects them:
| OTEL Attribute Prefix | Framework |
|---|---|
claude.managed_agents.* | managed_agents |
managed_agents.agent.id | Agent identifier |
managed_agents.agent.name | Agent display name |
API Endpoints¶
| Method | Path | Description |
|---|---|---|
POST | /managed-agents/sync | Primary: Pull sessions from Anthropic API |
POST | /managed-agents/discover | Discover agents from Anthropic API |
POST | /managed-agents/webhook | Receive pre-processed session data |
POST | /managed-agents/connections | Register a connection |
GET | /managed-agents/connections | List connections |
POST | /managed-agents/agents | Register an agent |
GET | /managed-agents/agents | List agents |
GET | /managed-agents/stream | SSE for real-time Pisama events |
Beta Notes¶
Claude Managed Agents is in public beta (managed-agents-2026-04-01). Multi-agent threads are in research preview. Pisama pins the beta version and gracefully handles unknown event types.