One trace for the whole agent team: Cogitator links workflows and workers
The October 21 release connects nested agents and distributed workers, with OpenTelemetry GenAI attributes and an explicit choice about recording prompt content.
Cogitator is making an agent team’s work followable in one trace. Its October 21 workflows 0.13.0 and worker 0.9.13 releases carry a tracing change that links nested agent runs, workflow nodes and distributed swarm workers instead of giving each run an unrelated trace.
For a builder debugging a researcher that calls another agent as a tool, the useful change is the parent-child chain: the inner run joins the tool’s trace, and workflow agents sit below their node spans. Distributed workers carry the same context. The merged implementation also addresses concurrent workflows accidentally nesting into one another and preserves trace context across a paused run and its resume.
Prompts are a separate recording choice
The spans add OpenTelemetry’s gen_ai.* attributes alongside Cogitator’s existing attributes. Model calls record provider, request and response metadata, token usage and streaming time to first chunk. W3C trace headers support carrying a parent context between components.
Prompt text, answers, tool definitions, arguments and results require captureContent: true on the runtime telemetry configuration or an observer. Observers that do not request content receive content-free spans. Studio requests content for its own observer, so developers inspecting prompts there should account for what their tools and models process. A shared trace is an observability feature, not a guarantee about tool safety.
A runtime for building agents
Cogitator is a self-hosted TypeScript framework, rather than a plugin installed into a coding harness. Its documented core installation is pnpm add @cogitator-ai/core zod; the project requires Node 22.12 or later. Local Ollama or vLLM and hosted providers are separate model backends. Hosted access needs the relevant credentials and provider charges; the MIT source license does not supply inference credits. Tool execution needs the permissions of the configured tools, with Docker or WASM sandboxing available through an optional package.
Upgraders should check integrations that parse identifiers: trace IDs become 32-character hexadecimal strings instead of trace_-prefixed values, and agent run spans become INTERNAL. Time-travel comparison and checkpoint lookup use a run ID, accepting a trace ID only when it identifies one recorded run. AITrending reviewed the release and implementation notes but did not run the framework or validate tracing overhead.