Fiddler sells software that records what AI agents and models did, judges whether they behaved correctly and safely, and blocks risky outputs in real time. Its buyers are AI engineering teams, often at banks, insurers, healthcare companies and government agencies. The risk staff who must answer for those systems also use it.

The input is a trace: a record of one agent run, including the instruction, each step, the tools called and the results. Teams send traces through OpenTelemetry, an open standard for collecting software data, with support for LangGraph, Amazon Bedrock, AWS Strands and Google ADK.

Fiddler nests this data from application to session, agent, trace and span, which is the smallest step. An engineer can start at a dashboard showing hallucination, toxicity, jailbreak or PII/PHI leak rates and click into one conversation to find the failing tool call.

Take a hypothetical insurer building a claims agent. Before launch, the team tests the agent against a "golden" dataset of past claims with known answers and compares prompt versions side by side. After launch, Fiddler scores live traffic and alerts on falling quality, and its guardrails can stop a response that would expose a policyholder's medical details.

Scoring runs on Fiddler's Centor models, which are small evaluators that check faithfulness, sensitive information and safety. Fiddler says they run inside the customer's environment without external LLM calls and enforce guardrails in under 80 milliseconds. Customers can also bring their own judge models.

Fiddler still monitors predictive models, such as credit scoring models, for drift. It deploys as SaaS, in a customer's virtual private cloud (VPC), on premises or in AWS GovCloud.

Enterprises pay Fiddler directly under three plans. The Free plan covers real-time guardrails. The Developer plan is usage-based at $0.002 per trace and adds observability, custom evaluators and SSO on Fiddler-hosted SaaS.

The Enterprise plan is negotiated. It adds VPC or on-premises deployment, scalability and a named customer success manager, and its contract terms are not public.

Per-trace pricing ties revenue to agent activity. For illustration, a hypothetical customer sending 10 million traces a month at the published Developer rate would incur $20,000 in trace charges, before any other applicable terms.

Fiddler's main costs are compute to score traces, storage, hosting and support. Smaller Centor evaluators offer a route to lower scoring costs than general-purpose LLM judges, but their compute, storage and development still have to be funded. Customer-hosted deployments can shift some infrastructure spending to the customer while adding implementation and support work.

If the acquisition closes, CrowdStrike will sell Fiddler inside its Falcon platform as Falcon Observe, with general availability planned for its fourth quarter of fiscal 2027. CrowdStrike has not announced Falcon Observe pricing.

Fiddler's most direct rivals are AI observability specialists such as Arize and LangChain's LangSmith, which also trace agent runs, run evaluations and monitor production quality. General monitoring vendors such as Datadog, along with the cloud providers' own tools, can add agent monitoring to bills customers already pay. The central contest is whether agent oversight stays an independent category or becomes a feature of the monitoring or security stack, and the pending deal places Fiddler in security.

Arize sells a similar loop of tracing, evaluation, monitoring and experimentation across predictive models and LLM agents. Fiddler’s case rests on its in-environment evaluators, deployment flexibility and work in regulated industries; both must turn broad monitoring capabilities into reliable, economical assessments of agent behavior.

LangSmith combines traces, evaluation datasets and testing with the developer ecosystem around LangChain and LangGraph. Fiddler must win a separate procurement decision against tools developers may already use, emphasizing governance and production oversight across frameworks rather than dependence on one agent-development stack.

Datadog's Agent Observability tracks LLM cost, latency, errors and quality. It also scans for sensitive data and prompt injection and links agent traces to the application monitoring engineers already use. Its advantage is that the data is already flowing into Datadog.

Fiddler counters with deeper evaluation and explanation. If the deal closes, its edge could shift to linking agent behavior with CrowdStrike's identity, endpoint and cloud security data rather than application performance data.

AWS, Microsoft and Google can package monitoring into the services where agents run. Fiddler partners with AWS inside SageMaker Studio, but it markets itself as working across Azure OpenAI, Bedrock, LangGraph and more than 100 other providers. CrowdStrike ownership would need to preserve that cross-platform compatibility.

Fiddler's core buyer has been the AI team monitoring models it built. Its larger expansion paths bring in buyers paid from security and compliance budgets. Both paths depend on execution and, in the security case, on the deal closing.

If the deal closes, Falcon Observe would reach the security teams at CrowdStrike customers, who investigate incidents and control access. CrowdStrike already treats agents as identities tied to a human and secures them at runtime. Fiddler would add a record of each agent's instructions, tool calls and results, along with evaluator judgments on whether it stayed on task.

That could let an analyst spot an agent with valid credentials taking an unsupported action, a failure that requires no attacker. Revenue would come through Falcon contracts rather than separate Fiddler sales, on terms not yet disclosed.

CrowdStrike's CEO said that when a regulator or board asks what an agent did, "the model did it" is not an answer, and he called Fiddler a "flight recorder" for agents. That points to a potential buyer in risk, compliance and internal audit, with a budget separate from engineering.

Fiddler’s VPC, on-premises and GovCloud options suit these buyers, with evaluator models that can run inside their environment and a history in model explainability. Many such buyers cannot send sensitive records to an outside evaluator. Selling audit reports and policy evidence to them is an analytical possibility, not an announced product line.

Pending deal: Closing still depends on customary conditions, and Falcon Observe integration is only planned. In the meantime, customers may delay standalone purchases until they see how the product is packaged and priced.

Lost neutrality: Fiddler markets itself as model- and cloud-agnostic, but CrowdStrike plans to sell it as part of Falcon. Enterprises standardized on rival security vendors could choose other observability tools instead.

Evaluation economics: Per-trace revenue helps only if scoring costs stay below the price charged, which depends on Centor models remaining cheaper than general-purpose LLM judges. Customers can get good ROI from cheap evaluation while Fiddler keeps little gross profit if bundled rivals push per-trace prices down.