NewAgent Observability is live.

Guides

AI observability for agentic engineering

A guide to monitoring your software when AI agents write, test, ship and debug it, with a directory of 90 monitoring and developer tools checked for what an agent can get out of each one.

If your team uses AI coding agents such as Claude Code or Cursor, five things need watching in new ways: the code they write, the pipelines that ship it, the systems that run it, the agents themselves, and the bill. This guide has one page for each. We call them stages because they follow the order of the work.

The five stages

Where agents change what you need to watch

Each stage page has four parts: what changes when an agent does the work, the numbers worth tracking (the guide calls them signals), three levels from doing it by hand to letting the system do it, and the tools that cover the stage. Use the levels to see where your team is and what the next step is.

  1. 1.Code, tests & review

    Agents writing code, generating tests and reviewing pull requests as one loop.

    Signals
    Review runs, not just pull requestsCost per review runTest coverage the agent added versus the coverage it claimed
    Read this stage
  2. 2.CI & release

    Pipelines running far more often, and releases compressing from weeks to hours.

    Signals
    Pipeline runs per merged changeTime from merge to productionFlake rate
    Read this stage
  3. 3.Production observability

    Infrastructure, APM, logs, traces, profiling and incident response, with an agent as the reader.

    Signals
    Result caps, and whether they truncate or errorWhether pagination exists at allWhat sampling and adaptive retention have already dropped
    Read this stage
  4. 4.Agent observability & evals

    Watching the agents themselves: hallucination, drift, eval scores and what a session actually did.

    Signals
    Eval score over time, per prompt versionTurn count and context growth per sessionTool-call failure and retry rates
    Read this stage
  5. 5.AI cost management

    Token cost, model routing and spend governance.

    Signals
    Cost per session, not cost per tokenRetry-driven token growthSpend by model, prompt version and team
    Read this stage

The vendor directory

Check what a tool's MCP server really returns before you trust the agent's answer

An MCP server is the connector that lets an AI coding agent query a tool such as Datadog or Grafana in plain English. Each directory page checks one vendor's server for what an agent can ask it and where the answers come back cut short with no warning. 63 of the 90 vendors have a working server of their own.