> ## Documentation Index > Fetch the complete documentation index at: https://docs.befailproof.ai/llms.txt > Use this file to discover all available pages before exploring further. # Custom agents > Wrap your agent in three `with` blocks and it starts recording. For an agent you wrote yourself, or a framework Failproof AI has no adapter for. You wrap the work you already do in `with` blocks. Each one records when it started, when it finished, and how long it took — you do not call anything else. ## Install ```bash theme={null} pip install failproofai-sdk ``` No extras, and no dependencies. ## Instrument ```python theme={null} import failproofai_sdk failproofai_sdk.configure(environment="production") with failproofai_sdk.session(): # one run with failproofai_sdk.agent("planner"): # one unit of work with failproofai_sdk.tool_call("search", input={"q": q}) as t: t.output = search(q) # one tool call ``` Read it top to bottom and it says what it means: | Wrap it in | To say | | ------------- | ---------------------------------------------------------------------- | | `session()` | These events belong to the same run | | `agent()` | Something is doing work — give it a name you would recognise in a list | | `tool_call()` | This is one tool, and here is what it returned | Nest them however your code nests. You never pass ids around: each block remembers what it is inside, so anything you record picks up the right run and the right agent on its own. `async with` works exactly the same. ## Check it arrived Run it once, then open **Observe → Sessions** and select your environment. Your run appears as a reconstructed trace. If nothing arrives, confirm the machine is connected with `failproofai config --status`. Every event method, what happens when your code raises, a worked example, threads and async, and instrumenting a framework with no adapter.