One bad tool call crashes the whole agent run

Your agent calls a tool, the tool throws, and the whole script dies instead of the agent trying something else.

Beginner to intermediateRead 6 minDoing: 30 to 45 minutes6 steps

▶ Start

What you'll have at the end

A tool wrapper that catches every failure, feeds the error back to the model, and caps retries at a fixed number

Before you start

You need

An agent loop that already picks a tool and calls its Python function directly with the arguments the model returned, with no error handling wrapped around that call yet.

Not covered

Deciding whether the model should keep trying automatically or wait for a person's go-ahead when the tool it is calling deletes, sends, or pays for something; this wrapper only decides between retrying a call and reporting that it failed.

Leans on

Build a wrapper that survives a bad call

Put every tool call behind one executor
Give every attempt a hard timeout
Sort every catch into retry or not
Retry the failure with backoff and a cap

Hand the model something it can act on

Report the failure instead of raising it
Confirm the model actually adapts

Where to go next

Log every tool call so you can find what broke a run

Once this wrapper is catching every failure, log each attempt so a bad run stays diagnosable instead of only survivable.

Should your agent ask before it deletes, sends, or pays?

If the tool this wrapper is retrying can delete, send, or pay, decide whether it needs a person's go-ahead before you let it fire again on its own.

Checked 15 Aug 2026

Part of the AI Agents cookbook