Key takeaways
Most failures come from messy inputs, not bad models.
Small, frequent evaluations beat big launches.
Visible logs turn sceptics into champions.
Looking back across millions of logged agent actions, a few patterns repeat. Very few of them are about model quality. Most are about the systems around the model.
Failures cluster at the edges
Agents rarely fail in the middle of a workflow. They fail when a tool times out, a record is missing or a policy changed last week. Retries and clear fallbacks prevent more incidents than prompt tuning does.
Small evaluations beat big launches
Teams that test against a small set of real examples before every change ship faster and break less. A hundred honest test cases will catch what a large demo never will.
Trust grows in public
When people can see what an agent did and why, adoption follows. Transparent logs turned sceptical reviewers into the strongest internal advocates we saw.
Cost follows retries
When we looked at where spend actually went, one pattern stood out. A small share of tasks, stuck in loops of retries and repeated tool calls, used a large share of the budget. Better models did not fix it. Clear stop rules did.
We now cap every task with a budget for steps and spend. An agent that hits the cap stops, writes down what it tried and asks for help.
What the numbers changed
Looking across customers, a few habits separated the teams that scaled from those that stayed stuck in pilot.
They reviewed failures every week, not every quarter
They kept test sets built from real cases
They gave each agent a single named owner
They expanded scope one task type at a time
Most trouble starts upstream
We expected model mistakes to lead our failure reports. They came third. The biggest group was stale or conflicting source data, followed by tools that returned something other than what their documentation promised.
That changed where we spend our time. Before tuning an agent, we now check the policies it reads and the systems it calls. A one line fix to a help article has repeatedly done more for quality than a week of prompt changes.
Boring beats clever
The most reliable agents in our data were not the most sophisticated. They did fewer things, used fewer tools and followed simpler instructions. Every extra tool added a new way to fail, and most added very little value.
If we had to summarise twenty three million actions in one line, it would be this: reliability is built from small, unexciting decisions, made consistently, by people who read the logs.
About the author
Theo Rask
Theo leads engineering at Falak and builds the permission, logging and evaluation systems behind every agent.
