Research project · Agent reliability
Agent Starter Kit
What is the smallest agent scaffold that is ready for production rather than ready to demo?
Reference buildThe Agent Starter Kit is a reference implementation built on CrewAI. It captures, in working code, the handful of practices our reliability work keeps coming back to - the ones almost every agent tutorial leaves out.
What we're investigating
Retries that know what to retry
Exponential backoff with jitter for transient failures, and an immediate stop for errors that will never succeed, such as a 404.
Failures that do not kill the run
A dead tool degrades one step. The agent is told what it can no longer rely on and adapts, instead of the whole run being thrown away.
Observable and testable by default
One structured event per agent action, tool call and task outcome, config-driven agents, fail-fast configuration, and a test suite that runs offline with no API key.
Approach
- 1Keep agents and tasks in configuration, not code.
- 2Take structured events from the framework's own event bus.
- 3Stub the model at the framework's documented extension point, so tests exercise the real code path.
- 4Write down what breaks in production and what to do about it.
Where it stands
The reference build exists and is used internally. It is not on sale.
Working on the same problem?
We'd like to hear from researchers, engineers and teams who have run into it too.
Get in touch