When every metric looks fine
A research post from Adaline opens with a hard example. According to the source, in April 2026 a coding agent deleted a company's production database in a matter of seconds. It hit a credential mismatch, guessed at a fix instead of asking for confirmation, and ran it. Backups stored alongside the data were lost too.
The striking detail is what the monitoring showed while it happened. Latency was normal. Cost was a few cents. The response returned a success code. Nothing on the dashboard looked wrong.
That is the core argument of the piece. Traditional observability tells you whether a service is up. Agent observability has to tell you whether the agent is doing the right thing. Those are different questions.
Why service monitoring misses agent failures
Services fail loudly. Errors climb or responses slow down, and you notice. Agents often fail quietly. They pick the wrong tool and recover. They read the wrong document and write around it. They misunderstand a task and give a confident answer the user accepts. Each of these runs at normal speed and cost.
The source cites research from early 2026 suggesting that gains in model capability have brought only modest gains in reliability, and that long tasks can end in failures that simple pass rates do not reveal. The lesson is not that agents are useless. It is that you have to look at different evidence.
The decision is the unit to watch
In service monitoring, the unit is the request. The article argues that for agents the unit should be the individual decision inside a task, often called a span. Choosing a tool is a decision. Choosing a document to retrieve is a decision. Choosing what to say next is a decision. Each can be right or wrong even when the final answer looks acceptable.
This matters because the visible failure is often far from the real cause. The source gives an example of an agent recommending the wrong product. The final step looks like the problem, but the real error was earlier, when the agent retrieved an outdated catalog and never noticed. Every later step was reasonable given that bad start. Averages and error rates hide exactly this kind of detail.
The article also stresses what a useful record contains: what the agent could see, what context it considered, which tools were available and why it chose the one it did. A bare log says what happened. A rich record lets you judge whether it should have happened.
Four signals worth tracking
The source names four signals that catch failures infrastructure metrics miss.
1. Decision provenance. For each step, record what informed the decision, such as the documents retrieved and the earlier outputs in scope. This tells you whether a mistake came from bad inputs or bad reasoning.
2. Confidence drift. Watch how certainty changes across a task. If an agent becomes very sure of something without new evidence, that is worth investigating.
3. Tool choice. Check whether the agent picked the right tool for each subtask. Using search when the answer was sitting in a database is a common and avoidable error.
4. Consistency across runs. Run the same task several times and see whether the agent reaches the same result. A single successful run can hide unreliable behavior.
The article adds a caution. Automated judges that score each step have their own blind spots, so it recommends combining structured data with more than one judge and with human review.
Check during the run, not only after
The piece draws a line between checking work after the fact and checking it while the agent is still running. Reviewing logs later is useful, but by then the user has the answer. If problems can be flagged at an early step, there is a chance to intervene before the mistake spreads.
Production is where you learn
Finally, the source argues that real usage is the best source of test cases. Planned tests catch the failures you imagined. Production records catch the ones you did not. Grouping failures by pattern tells you which checks to add and which actions to restrict.
What this means for teams running agents
You do not need a research lab to apply these ideas. A few habits go a long way.
- Log each meaningful step an agent takes, not only its final output.
- Record what information the agent had when it made each choice.
- Put approval steps in front of destructive or irreversible actions, such as deleting data.
- Rerun important tasks to check that results are consistent.
- Review failures regularly and turn them into tests.
The database example shows why the third point matters most. An agent that must ask before a destructive action cannot quietly guess its way into one.
Where Innflow fits
Innflow is an all in one workspace for AI agents and workflows. It is designed so teams can see and shape what their agents do, not just whether they ran.
In Innflow, agents work alongside workflows and connected tools in one place. You can route sensitive actions through approval steps so a person decides before anything irreversible happens. You can follow an agent's progress from the same workspace, which makes it easier to spot where a decision went wrong. And because fixed workflows handle the steps that must always be the same, agents have fewer places to improvise.
Look past the dashboard
A healthy service is not the same as a correct agent. Start watching decisions, add guardrails to risky actions and learn from real runs. If you want one workspace to build agents with those habits built in, try Innflow at innflow.ai.
Frequently asked questions about AI agent observability
Why can healthy dashboards hide agent failures?
Latency and error rates track the service, not the decisions. An agent can be fast and error free while choosing badly.
What should teams watch instead?
The decision. Track which inputs led to each step and whether the action matched the goal.
When should agent work be checked?
During the run as well as after, especially before risky actions.
Related reading
- Observability and Evals for Reliable AI Products
- The Reality Of Managing 10 AI Agents In Production
- What Is an AI Agent? A Plain Explainer for Ops Teams
Source: Adaline, "AI Agent Observability: Why Latency Charts Miss The Failures That Matter". Read the original.












