Agent ops: see, control and improve agents in production

An agent in production is a loop nobody watched: it chose tools, read data and spent money on its own. Agent ops is three jobs. See what each run did. Control what any run may do. Improve results per task, per dollar and per second. Most teams build the first, skip the second and argue about the third without data.

A trace shows one task as a tree of spans: the agent invocation at the root, its model calls and tool calls below it, and any subagents nested under the tool call that started them. A trajectory is the same run as an ordered list of steps, the messages and tool calls in sequence. You debug with traces and you score with trajectories.

The shared vocabulary comes from the OpenTelemetry semantic conventions for generative AI, now maintained in their own repository and still marked Development, so names can change. Agent work uses the operations invoke_agent, create_agent, invoke_workflow, plan and execute_tool, with attributes such as gen_ai.operation.name, gen_ai.agent.name, gen_ai.agent.id, gen_ai.provider.name and gen_ai.tool.name. A separate convention covers MCP calls.

Claude Code emits its own spans when tracing is on. Each prompt starts a claude_code.interaction span. Model calls are claude_code.llm_request children, tool calls are claude_code.tool children with blocked_on_user and execution spans beneath them, and a subagent's spans nest under the tool call that spawned it. When Claude Code talks to the Anthropic API directly, the W3C trace context travels into Bash commands through TRACEPARENT and onto model and HTTP MCP requests; behind a custom ANTHROPIC_BASE_URL proxy, set CLAUDE_CODE_PROPAGATE_TRACEPARENT=1 to keep it. The MCP 2026-07-28 revision documents the same traceparent key for MCP messages. An app that starts Claude Code through the Agent SDK can therefore see the agent's spans inside its own trace.

You need agent ops when any of these is true:

• Agents run unattended: routines, CI jobs, scheduled reports, background sessions.
• Other people rely on an agent's output, such as support replies or pull requests.
• Spend grows and nobody can say which team, model or tool drives it.
• You change prompts, models, tools or CLAUDE.md and want to know whether results got better.
• The agent touches regulated data and you must show who did what.

For one developer in interactive sessions, it is overkill. /usage shows session cost, prompt cache statistics and what counts against your plan, and /context shows what fills the window.

Pick tools by the job, not the logo:

If you build your own agent on the Claude Agent SDK, you get the same telemetry as Claude Code, because the SDK runs the Claude Code CLI underneath and passes your settings through.