Berserk for AI Ops
AI observability at scale, and the flexibility agents need.
Prompts, tool calls, and model reasoning are unstructured, high-cardinality, and text-heavy. They carry the operational signal that matters most, and they fit none of the fixed schemas observability was built on. Berserk is built for them, alongside your logs, metrics, and traces.
Not just one AI trace at a time
Because of Berserk's scalable tracing, you can aggregate and analyse agent behaviour across millions of runs, and join it with the rest of your telemetry, like metrics and logs, to optimize token usage, agent security, BI, anything really. And with Berserk's vector search and fast free-text search, prompts and responses aren't just text blobs, they're gold waiting to be mined.
Token-usage optimization
See which prompts, agents, models, and customers drive token spend, and what changed when it spiked. Tune the prompts and routes that actually move the bill, across every run.
Business intelligence
Quality, cost, and behaviour by customer, feature, and model over time, answered over the same data you debug with, drawing on free-text prompts and structured fields in a single query.
It's the wild west, and shadow AI is rampant
Organizations are scrambling to get their AI agents under control. Teams ship agents into production every week; people wire up their own. Prompts no one reviewed, tool calls no one logged, decisions no one can replay, agents running ungoverned, off in the corners of the org.
You can't secure, audit, or trust what you can't see. Getting AI under control starts with making it observable.
The fixed-schema assumption has broken
Observability wasn't built for AI. It assumed small, structured, low-cardinality inputs and predictable behaviour, and AI breaks every part of that assumption.
- ·Code
- ·Function arguments
- ·Configuration
- ·System prompt
- ·Tool instructions
- ·Context window
- ·Inputs are small
- ·Behaviour is predictable
- ·Meaning is structured
- ·Inputs are unstructured
- ·Behaviour is emergent
- ·Meaning is implicit
- ·Structured, low-cardinality fields
- ·Unstructured, high-cardinality text
- ·Deterministic code paths
- ·Model weights, billions of parameters
- ·Exact prompt phrasing
- ·Sampling temperature
Different shape, same need, both demand observability.
Observability needed a rewrite
Software produces more telemetry than ever, security events, error logs, application traces. Now add AI: prompts, tool calls, and model reasoning, generated on every request. Together they tell the story of your systems, but the plot gets lost in the volume.
AI makes it harder. Prompts and reasoning are large and text-heavy, and they capture the decisions driving your product, yet they fit none of the schemas traditional telemetry assumes. The signal is real; the shape is wrong.
What you need isn't a bolt-on system for AI logs. It's one system that correlates logs, metrics, traces, and AI output, and that doesn't depend on a human at a dashboard. The questions that matter in an agentic system are ones nobody thought to pre-build. Berserk gives you, and your agents, the data, the query language, and the latency to ask them.
What observable AI demands
Schemaless ingest
Prompts, tool calls, and reasoning don't share a shape, and it changes every release. Point OpenTelemetry at Berserk and ship, no migration when your agent stack shifts.
Efficient storage
AI telemetry is the new high-volume worst case. Storage efficient enough to keep every prompt and reasoning trace, not a 10% sample of the run that mattered.
Fast free-text search
The signal lives in large, text-heavy fields. Exploratory free-text search over prompts and completions, fast enough to iterate, so you find the run, not just count it.
Correlation with logs, metrics & traces
An agent run is also an application request. Join the model trace to the app trace, the error log, and the latency spike, one store, one query.
Where Berserk meets AI
Berserk wasn't built only for AI, but the way it was built lands exactly where AI telemetry needs it.
Free-text search for prompt analysis
Highly optimized search over the text-heavy fields where AI behaviour actually lives.
Semi-structured by design
OTEL spans, JSON, free text, and prompts, held side by side without a schema to negotiate.
Streaming results
Recency-first, streaming answers, so you (or your agent) act on signal without waiting for the full scan.
Case study
Governance you can prove
Berserk is working with Ethira to bring full back-testability and evidence gathering to Ethira's AI Agent Governance offering. Ethira maps and enforces what every agent can reach; Berserk keeps the durable, queryable record of what they actually did, every prompt, tool call, and decision, so policies can be back-tested across history and every governance call is backed by evidence.
ethira.devGetting started is trivial
Agent harnesses like Claude Code and Codex already emit OpenTelemetry. Point them at Berserk and ship, schemaless ingest means there is nothing to model first, and nothing to migrate when your agent stack changes next week.