Best Audit Logging Platforms for AI Agents
The real test for an agent's audit log is whether it records the calls that got blocked, not just the ones that ran.
Key takeaways
- -Most audit logging tools record what an agent executed, not what it tried and was stopped from doing, which hides the most useful governance signal.
- -AWS CloudTrail logs denied AWS API calls through its errorCode field, but only for AWS services, not the third-party or internal APIs most agents actually call.
- -Vanta and Drata automate SOC 2 evidence collection and continuously monitor controls; neither is a call-level audit log of what an agent asked an API to do.
- -Swytchcode's local audit log records both executed calls and policy-blocked calls, redacts sensitive values, and keeps 90 days of history on the Business plan.
Does your audit log know what your agent was stopped from doing?
Every vendor in this category will show you a log of API calls: timestamps, method names, response codes, maybe a trace ID. That part is solved. What most of these logs do not show you is the call an agent tried to make and never got to finish, because a policy, a permission check, or an allowlist stopped it first.
That gap matters more than it looks. A log of executed calls tells you what happened. A log that also includes refused calls tells you what your agent was about to do. If an agent tries to issue a $40,000 refund and a policy blocks it, the interesting event is the attempt, not the absence of a charge on a statement. Audit tooling built for humans clicking through a dashboard was rarely asked to capture that, because a human who gets an "access denied" error usually just stops. An agent tries again, with slightly different arguments, sometimes dozens of times a minute.
We went through the platforms teams actually reach for when they need an audit trail behind an AI agent: a cloud-native baseline, a general-purpose observability stack, a dedicated agent actions runtime, and the compliance automation tools that get lumped in with "audit logging" even though they do something different. For each one we asked the same question. If a call gets blocked, does the record say so, or does it just not show up?
How we scored these
We read each vendor's own documentation and pricing pages in November 2026, not marketing copy from a comparison site. Where a doc page did not say whether refused or denied actions are logged, we say exactly that, because a platform not describing a capability is not the same as a platform lacking it.
We weighted three things: whether blocked or denied actions appear in the log at all, how long records are kept without an add-on, and what pricing looks like at the volume an active agent produces, not a dashboard demo. A tool built for compliance evidence and a tool built for request-level audit logging are different products, and we tried not to flatten that difference just because both get called "audit logging" in a sales deck.
1. AWS CloudTrail
CloudTrail was never marketed as an audit log for AI agents, and it is the one most agent builders already have running underneath them. Any agent that calls AWS services, directly or through a tool wrapping the AWS SDK, is already generating CloudTrail events whether anyone configured anything or not.
Event history is free and covers the most recent 90 days of control-plane activity, and a trail delivers one copy of management events to S3 at no charge, with additional copies at $2.00 per 100,000 events. Data events cost $0.10 per 100,000 delivered. CloudTrail Lake adds queryable long-term storage, priced at $0.75 per GB for one-year extendable retention or $2.50 per GB (first 5 TB) for seven-year retention.
On the refused-call question, CloudTrail does better than most dedicated tools here, within its own boundary. AWS's documentation confirms an event record carries an errorCode and errorMessage field when a request fails, including authorization failures, so a denied call is a real record, not silence. The boundary is the problem: this only covers AWS APIs. An agent blocked trying to call Stripe, Salesforce, or an internal REST service generates nothing in CloudTrail, because CloudTrail only sees traffic against AWS's own control and data planes.
2. Arcade
Arcade positions itself as "the actions runtime for production AI agents," built around per-user, per-action authorization rather than a hosted catalog reached through someone else's OAuth screen. Its audit logging is on by default and records administrative actions automatically: who acted, through which surface (API, dashboard, CLI, SDK, or system), what resource changed, and when.
Arcade's audit-log documentation covers administrative changes and API access in detail, down to caller IP and a structured payload per entry, which is more granular than most competitors publish. What that page does not state is whether a denied or blocked tool call, as opposed to an administrative change, gets the same treatment. It names failures only as HTTP-level errors (401, 403, 429) and points to a separate tool-execution view for why a run failed, without describing how refused calls appear there. That is not a gap we're claiming Arcade has, only something not described in their public docs as of November 2026.
Pricing runs on tool calls and auth events rather than a flat tier: free for 2,000 tool calls plus 2,000 auth events, then Team at $25/month plus $0.10 per auth event and $0.01 per tool call. At real volume, the auth-event charge is the line to watch, since an agent that re-authorizes often will see it add up faster than the tool-call line.
3. OpenTelemetry-based observability setups
A large share of teams never buy an "audit logging product" at all. They instrument their agent's code with OpenTelemetry spans and traces, ship everything to whatever backend they already run (Datadog, Honeycomb, Grafana, or a self-hosted collector), and call that their audit trail.
This approach can, in principle, capture exactly what a dedicated product might miss. If the instrumentation emits a span for a policy rejection, that rejection sits in the trace data right next to the calls that succeeded. The catch is that this only happens because someone decided to code it that way. OpenTelemetry gives you the plumbing, not the judgment call about what counts as an auditable event. Nothing in the default instrumentation most frameworks ship distinguishes "the agent asked and the request went out" from "the agent asked and something stopped it," unless a developer adds that distinction by hand, and keeps adding it every time a new integration gets wired in.
Cost follows whatever backend you pick rather than a single vendor's price list. Shipping logs to Datadog, for comparison, runs $0.10 per uncompressed GB ingested plus an indexing fee starting around $1.06 per million events at 15-day retention on annual billing. Retention, redaction of sensitive fields, and whether a rejected call even gets a span are entirely up to whoever built the instrumentation, which is both the appeal and the risk of this option.
4. Datadog Audit Trail and Cloud SIEM
Datadog sells two products that get reached for here, and they answer different questions. Audit Trail captures audit events across Datadog's own configuration, access, and billing assets, the kind of thing a security team needs to prove who changed a dashboard permission or added an API key inside Datadog itself. Its product page does not describe capturing an AI agent's outbound API calls, refused or otherwise, because that was never the product's job. Events are searchable for up to 90 days, with longer archival available, though Datadog does not publish a price for Audit Trail.
Cloud SIEM could plausibly ingest agent traffic, but only if you pipe your agent's logs into it yourself, the same way you would with any SIEM. It is priced by volume analyzed, with security-log analysis running around $5 per million events on annual billing, separate from the $0.10-per-GB ingestion fee. Neither product ships with a built-in concept of "this agent call was blocked by policy." You would be building that distinction yourself, inside Datadog's UI instead of a generic collector.
5. Vanta
Vanta automates evidence collection for SOC 2, ISO 27001, and similar frameworks, and continuously monitors the controls an auditor will ask about. For a security team working toward a SOC 2 Type II audit, that layer is genuinely valuable, and it is the reason Vanta gets mentioned in the same breath as "audit logging."
It is not, however, a record of individual API calls an agent made or was blocked from making. Vanta's public materials describe automating evidence and monitoring control status, not capturing request-level traffic between an agent and the APIs it calls. If a questionnaire asks "do you log every API call your AI agent makes," Vanta is not the tool that answers that, even though it is the tool that will later ask you to prove you have one.
Vanta does not publish pricing. Buyer-reported ranges put Core plans (roughly 1 to 25 employees, one framework) around $10,000 to $15,000 a year, Growth around $24,000 to $45,000, and Scale, for larger organizations with multiple frameworks, at $50,000 to $120,000 or more, plus a separate fee paid to the audit firm. Treat these as third-party estimates, since Vanta keeps that number off its site as of November 2026.
6. Drata
Drata sits in almost the same spot as Vanta: continuous control monitoring and automated evidence collection aimed at the same SOC 2 and ISO 27001 audits, sold in tiers that scale with headcount and framework count. The two get compared constantly, for good reason, since the pitch and the product shape are close.
The same caveat applies. Drata's materials describe continuous monitoring of controls, not logging individual agent-to-API calls, allowed or blocked. If the need is proof that a specific refund call was stopped by a specific policy on a specific date, Drata's monitoring layer was not built to answer that at the call level, even though a well-run deployment will make a SOC 2 auditor happy about controls generally.
Drata also does not publish pricing. Buyer-reported figures put Core around $10,000 to $15,000 a year for small teams, Plus around $20,000 to $40,000, and Premium, for 250-plus employees with multi-entity support, from roughly $50,000 up to reported increases pushing top-tier contracts toward $100,000 a year in 2026. These are estimates from buyers and resellers, not Drata's own published number.
Side-by-side comparison
| Platform | Logs refused/blocked actions? | Retention | Pricing (as of Nov 2026) |
|---|---|---|---|
| AWS CloudTrail | Yes, for denied AWS API calls via errorCode/errorMessage; not for non-AWS APIs | 90 days free event history; CloudTrail Lake offers 1-year extendable or 7-year retention | Management events free (1 copy); data events $0.10/100k events; Lake from $0.75/GB ingested |
| Arcade | Not described in public docs for denied/blocked tool calls; admin actions and API access are logged by default | Not published | Free: 2,000 tool calls + 2,000 auth events; Team $25/mo + $0.10/auth event + $0.01/tool call |
| OpenTelemetry-based DIY setup | Only if a developer instruments it; no default distinction between executed and refused calls | Depends on the chosen backend | Backend-dependent, e.g. Datadog logs at $0.10/GB ingested plus indexing fees |
| Datadog Audit Trail / Cloud SIEM | Not described in public docs for agent/API-level blocked calls; Audit Trail covers Datadog's own config/access changes | Audit Trail: ~90 days searchable, longer via archive | Audit Trail price not published; Cloud SIEM analysis ~$5/million events (annual) |
| Vanta | Not built for call-level logging; continuously monitors compliance controls instead | Not applicable (compliance evidence, not a call log) | ~$10,000–$120,000+/year by tier (buyer-reported estimate, not published) |
| Drata | Not built for call-level logging; continuously monitors compliance controls instead | Not applicable (compliance evidence, not a call log) | ~$10,000–$120,000+/year by tier (buyer-reported estimate, not published) |
Where Swytchcode fits (and where it doesn't)
Swytchcode is not a SIEM, and it is not a SOC 2 evidence tool. It is the layer that sits between an agent and the APIs it calls, and every call passing through it goes through the same eight-stage pipeline: allowlist check, schema validation, policy evaluation, credential injection, execution, retry, response validation, and audit. The last stage is the relevant one here. Every call and every policy decision gets written to a local audit log, so a refund blocked for exceeding a spend policy shows up with the same certainty as a refund that went through.
That is the direct answer to the question this post is built around. swy audit network shows what actually ran. swy audit policy shows what got stopped and why. Having both commands, pulling from the same local record, is what the catalogs and compliance platforms above do not offer as a built-in feature: most either show only executed traffic, or they're not answering this question at all. We cover this in more detail in how to log every API call an AI agent makes.
The honesty required on our side: this is a local audit trail, not a compliance program. Retention tops out at 90 days on Business (7 days on Free, 30 on Pro, custom on Enterprise), nowhere near CloudTrail Lake's seven-year option, and sensitive values are redacted before anything is written, by design, which keeps a credential out of a log file but means the raw payload isn't there later for a different kind of investigation. A SOC 2 report still needs Vanta, Drata, or an auditor's own framework doing the evidence-collection work; Swytchcode gives you the underlying record of what an agent's calls actually did, one input into that evidence, not a replacement for it. The policy side, including how argument-level rules decide what gets blocked in the first place, is covered in API compliance for AI agents and the execution pipeline guide.
The practical split: use a catalog like Arcade when per-user authorization against thousands of SaaS tools is the job, use Vanta or Drata when the deliverable is a compliance report an outside auditor signs off on, use CloudTrail as a baseline if the agent lives mostly inside AWS, and reach for something like Swytchcode when the question is narrower and sharper: for this specific call, to this specific API, did it run, and if not, why not. See API observability for AI agents and policy-as-code for AI agents for more on turning guardrails into code rather than instructions.
FAQ
Does an audit log need to include blocked or denied calls to be useful?
For governance purposes, yes. A log of only executed calls tells an investigator what happened, but not what an agent attempted and was stopped from doing, often the more important signal when something looks wrong. If a prompt gets manipulated or a tool misfires, the first evidence usually shows up as a refused call, not a successful one.
What's the difference between API observability and API audit logging?
API observability, the kind OpenTelemetry and Datadog provide, is about understanding system behavior: latency, error rates, traces across services. API audit logging is narrower: a durable, often immutable record of who or what did what, when, and whether it was permitted. The two overlap, since a trace can double as part of an audit trail, but observability tools aren't built with audit-grade retention or tamper-resistance as the default goal, while a purpose-built audit log is.
How long should an AI agent's audit trail be retained?
It depends on what the record needs to survive for. A security team debugging last week's incident is fine with 30 to 90 days. A compliance program answering to SOC 2 or a multi-year regulatory requirement needs something closer to CloudTrail Lake's seven-year option, or an export pipeline that moves records out of a shorter-retention tool before they age out. Match the retention window to the slowest-moving requirement you have, not the fastest.
More content
Best Agent Frameworks for Production Use
We rank six agent frameworks used in production as of November 2026 and show why none of them handle API permissions, retries, or audit by design.
Best API Catalogs and Registries for AI Agents
We rank Backstage, Postman, Stoplight, SwaggerHub and the agent-native catalogs on one question: does the schema stay current at call time?
Best API Platforms for Fintech AI Agents
We rank Stripe, Modern Treasury, Plaid, Adyen, Increase, and Column on the two things that actually matter for fintech agents.
