Research project · Agent reliability
Docket
When an automation fails, can the cause be attributed automatically - and proven?
In progressDocket studies fault attribution in production automations, using n8n execution histories as its testbed. Most tooling can tell you a run failed. Far less can tell you why, and almost none notices the run that finished green and quietly did nothing.
What we're investigating
Silent failures
Runs that report success but wrote nothing, or returned the right shape with the wrong values. We check what a run actually produced against what that workflow normally produces.
Cause classification
Sorting each failure into a rejected or rotated credential, a change upstream, or the workflow's own logic, with the execution record, the error and the previous run attached as evidence.
Read-only by design
How far attribution can go using only read access to execution history, with no agent installed and no changes to the workflows being studied.
Approach
- 1Read execution and workflow history through a read-only API key.
- 2Build a baseline of normal output for each workflow.
- 3Flag runs that deviate from it, whatever status they reported.
- 4Attribute a cause and keep the evidence alongside it.
Where it stands
In active development. Early work focuses on silent-failure detection and the attribution model itself.
Working on the same problem?
We'd like to hear from researchers, engineers and teams who have run into it too.
Get in touch