Skip to content
be-human.ai
← Back to the journal

Personalized AI infrastructure

AI Agent Observability: Traces, Cost Records and Failure Signals to Track

What a company needs to see what its AI agents do: task traces, cost records per task, and failure signals with owners and review dates. A practical guide.

Be Human typographic cover with the title: AI Agent Observability: Traces, Cost Records and Failure Signals to Track
Conceptual illustration created with AI · Be Human

AI agent observability is the practice of recording what an agent did, which records and tools it used, what it cost and where it failed, so that people can judge its behavior and change it safely. For a company, that means a trace of every step an agent takes, a cost record attached to every task, and a short list of failure signals that a named person reviews on a fixed schedule. Stronger models make this more necessary, not less, because a more capable agent can reach a correct answer through a path nobody expected.

Key takeaways

  • Observability is the evidence layer that shows what an agent actually did, which is different from what its instructions say it should do.
  • A useful trace links every model call, tool call, memory lookup and approval decision to one task identifier.
  • Cost records should be attached to units of work, such as a resolved request, so that a rise in spending can be judged against output.
  • Failure signals need thresholds, owners and a review date, or they become logs that nobody reads.
  • Set up observability in the first engagement, before any model upgrade, so that you have a baseline to compare against.

Why stronger models are harder to read

A weaker model often fails in obvious ways: it produces a clumsy answer, misreads a field or refuses a simple task. A stronger model fails less often in those ways, but its failures become harder to spot. It may complete a multi-step task correctly while reading three documents where the old version read one, calling a tool that the old version never touched, or asking for approval less often because it judges a step to be routine. The final output can look right while the path behind it has changed in ways that matter for cost, privacy and permissions.

This is why reviewing outputs alone is not enough. An operations leader needs to see the path: which records were read, which tools were called, which permissions were used, and where a person was asked to decide. Without that path, a model change can appear to improve results while quietly increasing risk. Observability turns those hidden changes into records that can be compared, questioned and reversed.

What a trace should record

A trace is the ordered history of one task from start to finish. You do not need a complex system to begin, but you do need consistent fields. For each task, the trace should include:

  • The task identifier, the requesting person or system, and the time the task started.
  • Each model call, with the model version, the instruction version and the time it ran.
  • Each tool call, with its parameters, its result status and the account or permission it used.
  • Each memory item retrieved, with its source and the date it was stored or last updated.
  • Each approval request, the decision taken and the person or rule that made it.
  • The final output, and whether a person accepted it, edited it or rejected it.

The edit and rejection fields are often the most useful. They show where the agent's output needs human judgment and where it is reliable enough to stand alone. The OpenTelemetry project describes traces as connected records of work across steps, and its semantic conventions for generative AI define common names for model and agent activity. Using shared names makes it easier to move records between tools later, which matters if you change platforms. For background on how these concepts fit together, the [OpenTelemetry observability primer](https://opentelemetry.io/docs/concepts/observability-primer/) is a useful starting point for a technical team.

How to track cost per task

Agent costs are rarely a single line. They come from model usage, tool calls, retries, storage of memory, and the human time spent reviewing results. A monthly invoice from a model provider tells you the total, but not which work produced it. The FinOps Foundation frames this as unit economics: relate cost to a unit of business output, so that spending can be judged against what it bought. For an agent, a practical unit might be a resolved customer request, a prepared quote or a reconciled record.

To make this work, attach cost to the task identifier. Record the usage of each model call and tool call inside the trace, then total them per task. Pay special attention to retries and loops. An agent that repeats a failed step three times may appear to finish successfully while costing several times the expected amount. Review the cost per unit of work each week during early operation, and note any change that follows a model, prompt or tool change. A cost rise is not automatically a problem, but it always needs an explanation that someone can give.

Which failure signals to watch

A failure signal is an observable pattern that suggests the agent is drifting from acceptable behavior. Signals are more useful than general quality scores because each one points to a specific place to look. Useful signals include:

  • Tool calls that fail, repeat or return empty results more often than the baseline you recorded in the first weeks.
  • Approval requests that rise sharply, or approvals that are granted so quickly that they suggest nobody is reading them.
  • Outputs that people edit heavily before sending, which often means the agent is wrong in a consistent way.
  • Memory lookups that return outdated records, or records that belong to a different customer, location or project.
  • Attempts to use a permission the agent was never granted, even when the attempt is blocked.
  • Content retrieved from documents, emails or web pages that contains instructions trying to change the agent's behavior. The OWASP Top 10 for LLM Applications lists prompt injection as a leading risk category, so this signal deserves its own owner.

Set each threshold from your own baseline, not from a general benchmark. Assign one owner per signal, a review day, and a clear action: investigate, pause the affected tool, or raise an incident. Signals with no owner tend to be reviewed by no one.

A checklist for the first month

Use this sequence when you put observability in place for a single agent with one defined role:

  • Write the one task the agent performs and describe what a good result looks like in plain words.
  • Define the trace fields listed above and store them under one task identifier.
  • Choose a cost unit, such as cost per resolved request, and record it with every task.
  • Select five failure signals, assign an owner to each, and fix a weekly review day.
  • Record a baseline over the first weeks, before any model, prompt or tool change.
  • Require that every change be compared with the baseline, and keep the previous configuration available so you can return to it.

Illustrative scenario: a quote assistant after a model upgrade

The following scenario is illustrative and does not describe a real company. A regional distributor uses an agent to prepare quotes for its sales team. After the model is upgraded, the sales operations lead notices that quotes are still accurate, but the team has not reviewed the trace data. The traces show that the agent now reads two pricing sheets per quote instead of one, and that one of those sheets is an older version. The cost record shows a higher cost for each prepared quote. The failure signals show that the outdated sheet was used on several occasions.

Because the trace named the sheets, the lead could restrict the memory source to the current pricing file, adjust the approval rule for quotes above a set value, and rerun the previous week's tasks to confirm the fix. Without traces, the output alone would have looked fine until a customer noticed a wrong price. The point of the scenario is not that the upgrade was bad. It is that the company could see what changed and decide what to do about it.

How Be Human runs this in a managed engagement

Be Human's managed approach treats observability as part of the agent's design rather than an add-on. During the first engagement, the trace fields, cost unit and failure signals are defined alongside the agent's role, its memory, its tools and its approval boundaries, and they are kept under the client's control. Recurring evaluations and permission reviews use these records to check whether the agent still behaves as intended. When a signal crosses its threshold, the managed process routes it to a defined incident response, and any change to the agent follows a written request and approval step, as described in How to Request and Approve Changes to a Managed AI Agent.

Where to start

Observability is easiest to add when the first task is small, well defined and important enough to review. A narrow task gives you a clear trace, a clean cost unit and a short list of signals that matter. Teams that try to observe every process at once usually record too much and review none of it. Begin with the one task where a silent error would cost the most, and expand from there as the review routine works. If you are unsure which task should be the first one with traces, the free Be Human AI diagnostic will help you identify it.

Frequently asked questions

Is AI agent observability the same as application logging? Logging records events, but agent observability links those events into the full history of one task, including model calls, tool use, memory lookups and approvals. Logs can exist without that link, which makes them hard to use when you need to explain a single outcome. The useful question is whether you can reconstruct why an agent produced a particular result.

Do we need specialized tools to start observing an agent? Not necessarily. You can begin with structured records keyed to a task identifier, stored in a system your team already controls, and reviewed on a fixed schedule. Specialized tools become useful as volume grows or as you need dashboards across several agents. Shared naming conventions, such as the OpenTelemetry generative AI conventions, make a later move easier.

How much detail should we keep, given privacy concerns? Keep enough to reconstruct decisions and no more than your policy and applicable law allow. Where a trace includes personal data, limit the fields, set a retention period and restrict who can read them. Companies operating in Brazil should check their obligations with the national data protection authority, the ANPD, at [its official portal](https://www.gov.br/anpd/pt-br), and companies in regulated sectors should align retention with their own compliance rules.

Who in the company should read agent traces? The process owner should review the traces for their area, and whoever approves permissions should see the access and approval records. Most staff do not need raw traces, and giving them broad access increases privacy exposure without improving decisions. Summaries of cost and failure signals are usually enough for executives, while detailed records stay with the people accountable for the agent.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →