Agenticx Agent Obs

AgenticX Observability Assistant (agenticx-agent-obs)

Guide for connecting AI agents to the AgenticX Runtime Observability center (/ops/logs: Overview, Traces, Observations, Sessions) and diagnosing trace errors or latency bottlenecks.

Core Capabilities

1. Desktop & Local Runtime Auto-Onboarding: Configure ~/.agenticx/config.yaml so Near desktop sessions automatically report run, model, tool, and subagent spans (including reasoning thought previews) to the enterprise observability endpoint.
2. Multi-Framework SDK Instrumentation: Instrument Python (agenticx, openai, langchain, langgraph), TypeScript/Node.js, or Go services to report structured trace spans.
3. Interactive Trace & Session Diagnosis: Query recent traces, inspect failed tool / model / subagent observations, and summarize latency P50/P90 or token spikes using the bundled diagnostic script.

Workflow 1 — Connect Near Desktop to Enterprise Observability

When the user wants their local Near desktop runs to appear in the Enterprise Admin Console (运行观测):

1. Inspect ~/.agenticx/config.yaml under the enterprise: section.
2. Ensure the following keys are present (ask the user for base_url and token only if missing):

   yamlCopy code
   enterprise:
     enabled: true
     base_url: "<ENTERPRISE_PORTAL_URL>"
     user_email: "<USER_ACCOUNT>"
   

   (The personal access credential is configured in Near Desktop Settings → Enterprise Connection.)
3. Verify connectivity using the bundled diagnostic script (see Workflow 3).
4. Explain privacy guarantees:
   • Uploaded spans include trace_id, step_no, step_kind (run | model | tool | subagent), status, model, provider, token counts, duration_ms, and 500-char clipped previews of user input, assistant reply, and model reasoning (metadata.thought).
   • Tool arguments and raw tool outputs are never uploaded.

Workflow 2 — Instrument External Agent Code (Python / LangChain / OpenAI / TS)

When the user wants to report traces from their own application code into AgenticX Observability:

1. Read references/ingest-guide.md in this skill directory for complete payload schemas and copy-pasteable code templates.
2. Choose the matching integration path:
   • Desktop Ingest (POST /api/desktop/agent-traces): Uses the enterprise connection header against the Web Portal. Best for developer machines, CLI agents, and custom Python/LangGraph runners.
   • Gateway Proxy (:8088/v1/chat/completions): Route OpenAI-compatible SDK calls through the AgenticX Go Gateway with X-Trace-Id, X-Trace-Step, and X-Session-Id headers for automatic span + token + cost capture.
3. Keep span hierarchies linked by setting metadata.parent_step_no on child model, tool, and subagent spans pointing to the root run step (or active model step for tool calls).

Workflow 3 — Diagnose Failed Traces, Slow Spans, or Token Spikes

When the user asks "Why did my last agent run fail?", "Which tool is slowing down the session?", or "Show me today's token usage and errors":

1. Run the bundled diagnostic script located in this skill's scripts/diagnose_traces.py:

   bashCopy code
   python3 ~/.agenticx/skills/registry/agenticx-agent-obs/scripts/diagnose_traces.py --mode summary
   

   (If running from the built-in path, use the directory where this SKILL.md resides.)
2. Supported flags:
   • --mode summary: Prints 24h KPI overview, error counts by category (tool / model / subagent / run), P50/P90 latency, and top models.
   • --mode errors: Lists recent failed traces and the exact observation step (#step_no, tool/model name, error message) that bubbled up to the trace.
   • --trace-id <id>: Prints the full parent-child call tree (run → model → thought / tool) and input/output previews for a single trace.
3. Summarize the root cause clearly for the user and provide actionable next steps (e.g., fixing a broken MCP tool command, adjusting model timeout, or narrowing prompt context).