AI Agent Observability: What to Log, Trace and Alert On
What to log, trace and alert on for AI agents: decisions, tool calls, costs and approvals a business needs to see to keep oversight as agent work grows.

AI agent observability is the practice of recording enough about each agent run to reconstruct what it saw, which tools it called, what it decided and what it changed, then alerting on the signals that show it has left its approved boundaries. For a business, the goal is not a large pile of logs but a short, reliable record that an operations leader can read and act on. This article covers what to log, how to trace a run from request to outcome, which alerts deserve attention, and how to keep the record readable for the people who own the process.
Key takeaways
- Log decisions, approvals and side effects, not only prompts and final answers.
- Give every run one identifier that follows it through tool calls, memory reads, handoffs and escalations.
- Alert on boundary crossings, cost and time anomalies and quality drift, not on every error.
- Write run summaries that the process owner can read without asking an engineer.
- Pair logs with a small, regular review of outputs, because a log shows what happened but not whether it was good.
Why agent observability is different from application logging
Traditional application logs record requests, responses and exceptions. An agent adds steps that depend on context: it retrieves documents, chooses between tools, plans a sequence and sometimes waits for a person. A run can finish with no technical error and still do the wrong thing, such as updating a record it should only have read or answering from an outdated source. OpenTelemetry describes observability through traces, metrics and logs, which together let you follow a request across components and see where time and decisions went. The [OpenTelemetry primer](https://opentelemetry.io/docs/concepts/observability-primer/) is a useful reference for those three signals, and its generative AI conventions extend them to model calls and agent steps.
The practical consequence is that you need to design the record before the first run. Adding visibility after an incident usually means you cannot reconstruct the decision that caused it.
What should every agent run record?
For each run, the record should answer five questions: who or what started it, what context it used, which actions it took, what it changed and how it ended. In practice that becomes a short list of fields:
- A run identifier, a parent request identifier, the agent role and the version of its instructions.
- The trigger and the input received, with personal data masked or minimized where the purpose does not need it.
- The context sources used, recorded as identifiers of documents, memory entries or records rather than full copies where possible.
- The model, provider, key settings and prompt version used for each model call.
- Each tool call with its name, arguments (redacted where sensitive), result status, duration, and whether it was a read or a write.
- The chosen action, the rule or approval that allowed it, and any alternative the agent recorded.
- Every human touchpoint: approval requested, granted, denied, or an escalation with its reason.
- The outcome: completed, handed off, blocked or failed, with the final state of anything it changed.
- Token use and cost for the run and for each step.
Do not record everything by default. Raw prompts and customer details carry their own retention and access obligations. The [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) frames this as a governance question, not only a technical one: decide what evidence you need to manage risk, then keep that evidence with a defined owner and a retention period.
How do you trace one run from request to outcome?
A trace is the chain of steps that belongs to one run. Build it with a simple sequence:
1. Assign the run identifier at intake and pass it into every tool call, message and memory write. 2. Record the start and end of each step with a parent link to the step that triggered it, so the order and nesting are visible. 3. Attach decisions as structured events with fixed fields, not only as free text inside a long message. 4. For any write, store the state before and after the change, so a reversal is possible and the change is explainable. 5. Link every escalation to the same run identifier, so the person who picks it up sees the full chain. 6. Test the design once: choose a completed run, reconstruct what happened using only the trace, and check that you did not need to ask the builder anything.
If step six fails, the trace is missing context that a reviewer will need during an incident.
A trace that only an engineer can read is a diagnostic tool. A trace that the process owner can read is an oversight tool.
Which alerts matter, and which are noise?
Alerts should map to decisions someone has to make. Too many alerts teach people to ignore them, and too few leave a looping or drifting agent running unseen. A simple decision guide helps sort them:
- Interrupt someone now: an action outside the approved permissions was attempted, a write failed partway through a sequence, or a tool credential was rejected repeatedly.
- Review the same day: cost per run moved well outside the agent's normal range, a step exceeded its time budget, or retries concentrated on one tool.
- Review weekly: sampled outputs changed in accuracy or tone, the escalation rate shifted, or an expected document source stopped appearing in retrieved context.
Cost signals deserve their own rules because a looping agent can consume budget without producing errors. The cost spike alerts guide covers how to set thresholds that catch runaway loops early. For handoffs, the escalation pattern for agent failures explains how to send the decision and its context to a person rather than the full mess of the run.
How do you keep the record readable for the process owner?
The person accountable for a process, such as a finance lead or support manager, should not need to parse raw traces. Generate a short plain-language summary for each run that states what was asked, what the agent did, what changed and whether anything needs a decision. Keep the summary generated from the structured fields, so it cannot drift away from the record.
Illustrative scenario: an intake agent for supplier invoices reads an email, matches it to a purchase order and drafts a payment entry for approval. The summary for one run might say that the invoice matched the order, that the agent drafted an entry without posting it, that the amount differed from the order by a small margin, and that the entry is waiting for approval. The owner can decide in under a minute, and the full trace is one click away if the decision is disputed.
How do logs and quality reviews work together?
Logs answer what happened. They do not answer whether the result was good. A run can pass every technical check and still produce a weak answer, a wrong category or a tone that does not fit the customer. Pair the trace with a recurring sample of completed runs, scored against a small set of cases the business already accepts as correct. Start with a handful of outputs each week, then adjust the sample size based on risk and on what the reviews find. The review results should be stored alongside the run identifiers so that a quality dip can be traced back to a specific version, tool or source.
Quiet degradation is the hardest pattern to see in logs alone, because nothing fails. Compare reviewed quality over time and watch for shifts in escalation rates and retrieval patterns. These are the signals that usually reveal drift before a customer or colleague reports it.
How Be Human approaches agent observability
Be Human designs and operates AI agents under human control, and observability is defined at the start of the work rather than added later. A first engagement defines the one useful role, the memory and tools it needs, the approval boundaries and the success criterion, and the trace fields and alerts follow from those decisions. Managed operations then cover monitoring, recurring evaluations, permission review, integration maintenance and incident response, so the record keeps matching the work as tools and permissions change. Service is remote, in English and Portuguese.
If you want to see which task in your business is the right first one to hand to an agent, and what it would need to log from day one, take the free Be Human AI diagnostic.
Frequently asked questions
What is AI agent observability? It is the ability to reconstruct what an agent did in each run, including the context it used, the tools it called, the decisions it made, the approvals it needed and the final state it left behind. It combines structured logs, traces that link steps together and alerts tied to the boundaries the business cares about. The aim is to let a person understand and correct the agent's work without guessing.
Do I need to store every prompt and response? Usually not in full. Store identifiers for context sources, structured fields for decisions and tool calls, and only the text needed to explain an outcome. Apply retention limits and access controls to anything that contains personal or confidential data, and mask it where the purpose does not require it.
How long should agent logs be kept? The answer depends on your legal, contractual and internal review obligations, so confirm it with the people responsible for compliance and records. Keep enough history to investigate an incident and to compare quality across agent versions. Document the retention period with the owner of the process, so it is applied consistently.
How do I know whether my observability is enough? Test it. Pick a completed run, rebuild what happened using only the trace and the run summary, and see whether the process owner can decide what to do next. If a reviewer still needs to ask the builder questions, add the missing fields or events, then test again on another run.
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.


