When a traditional web service misbehaves at 3 a.m., you open your APM dashboard. When an autonomous agent misbehaves, too often you open a wall of unstructured log text and start guessing.
Agents break the assumptions APM tools were built on: their "requests" are reasoning steps, tool invocations, and nested sub-agent work. That's why MeshCtx — an open-core (AGPLv3) agent platform — ships a standardized telemetry layer out of the box. Here's the surface, and how it fits together.
The four pieces
1. Structured tracing (OpenTelemetry-style). Every run produces compliant trace/span IDs, with nested spans attributing child work — a tool call, a spawned sub-task — back to the parent step. A 40-step agent run reads like a tree, not soup. The format is backward compatible with what you already parse.
2. Local-first JSONL. Telemetry lands as local JSONL files with automatic rotation at 2 MB and a ring-buffer cap. Nothing leaves the machine unless you send it somewhere. Tail it with jq, pipe it into your own pipeline, or keep it purely as forensic evidence of what the agent did while you slept.
3. Three HTTP endpoints. GET /events lists what happened, GET /stats aggregates it, and POST /record pushes your own events into the same stream — so plugin authors and your own scripts emit telemetry that lands in the same place. Records are tied to their task ("task-card level observability"), which is what makes them queryable per job rather than per process.
4. Optional OTLP export. Run a collector? Point MeshCtx at it and spans flow into whatever backend you already run:
MESHCTX_OTLP_ENDPOINT=http://collector:4318
Jaeger, Tempo, your vendor of choice — no lock-in, no proprietary sink. Available in Personal, Team and Enterprise editions; the feature landed in v3.123+ and the current release line is v3.131.x. The full reference is on the telemetry docs page.
Why telemetry is a first-class citizen here
MeshCtx's broader pitch is auditable, self-adaptive agents, and the three mechanisms cover three different questions:
- Governance (RBAC) answers what the agent is allowed to do.
- The hashed audit chain answers what it actually did — tamper-evident, ordered.
- Telemetry answers how the work performed, step by step: latencies, nesting, failure points.
If you've ever reconstructed an agent failure from chat transcripts, you know why you want all three.
Small aside: 11 languages, right-to-left included
The UI and docs ship in 11 languages (English, Chinese, Japanese, Korean, French, German, Spanish, Italian, Arabic, Hebrew, Russian). The recent v3.131.x line put real work into i18n coverage — including RTL behavior tests for Hebrew and Arabic, which is rarer than it should be.
Try it
MeshCtx is open core under AGPLv3, with downloads for Windows, macOS (Apple Silicon / Intel) and portable Linux on the releases page. Docs live at meshctx.com.
If you're building agents and your current debugging story is "scroll and pray," a standardized telemetry layer is the cheapest upgrade you can make.
Top comments (0)