AGENT-NATIVE TELEMETRY

Telemetry built for agents,
not dashboards.

OTEL ingest. SQL query. Built on object storage.
Keep every high-cardinality dimension forever, with predictable storage and query pricing.

No spam · Talk to the builders

BOOK A CALL

Built for AI infrastructure teams
investigating real production systems.

AGENT SESSIONinvestigator-07 / LIVE

agent.investigation

“Find refusals that never emitted a policy or error tag.”

SELECT trace_id, span_name,
  semantic_match(
    "gen_ai.output.messages",
    'sorry, I cannot help you'
  ) AS refusal_score
FROM spans
WHERE service = 'support-agent'
  AND ts > now() - interval '30 minutes'
ORDER BY refusal_score DESC
LIMIT 50;

trace_idspan_namerefusal

8fd1...a02csupport-agent.answer0.97

f21b...9e17support-agent.answer0.91

↳ finding127 likely refusals in the last 30 minutes; 63 emitted no policy or error tag. confidence 0.96

01 / OTEL-NATIVE02 / INFINITE RETENTION03 / STANDARD SQL04 / PREDICTABLE PRICING05 / ZERO DASHBOARDS

[ THE THESIS ]

Observability changed users.

The existing observability stack is built for humans. Agents don’t need another pane of glass. They need cheap, complete context and a language they can compose.

SQL, not UI

Agents already write queries, test hypotheses, and iterate. Give them a stable, expressive interface instead of teaching them where to click.

interface := query

Infinite retention

Keep all traces, logs, and metrics on low-cost blob storage with no retention window. Use what matters when the incident actually happens.

retention_days := ∞

Cardinality is signal

Keep all tenant, prompt, region, model, commit, tool call without aggregation. Agents need all of these to investigate the root cause.

dimensions := unbounded

[ THE STACK ]

Three primitives.
No theater.

The economics come from a deliberately boring data path. Send standard telemetry, retain it indefinitely, and let agents query raw history with tools every model understands.

INGEST

OTEL in

Traces · logs · metrics

grpc://ingest.tracestore.dev

STORE

Blobs at rest

Columnar · durable · cheap

s3://telemetry/raw/

QUERY

SQL out

Flexible · composable · agentic

SELECT * FROM spans

PRIMARY USER

Agent investigates

Queries complete history and tests hypotheses against raw telemetry.

ASK → QUERY → EVIDENCE → ANSWER

No proprietary agent SDK. No new query DSL. No instrumentation rewrite.

[ INVESTIGATION LOOP ]

Let agents ask the expensive questions.

Incidents don’t follow dashboard layouts. Agents move across signals, refine their scope, and join context that no one thought to pre-aggregate.

01
Form a hypothesis
Start from the alert, user report, or failed evaluation.
02
Query raw context
Slice by any attribute. Join across telemetry types.
03
Converge on cause
Run the next query, preserve evidence, report confidence.

THINK

Latency is isolated to one tenant. I’ll correlate the slow traces with deployment and infrastructure attributes.

QUERY

SELECT deploy_sha, k8s_node, p95(duration_ms) as latency_ms
FROM spans
WHERE tenant_id = '7f1c'
GROUP BY deploy_sha, k8s_node
ORDER BY latency_ms DESC
LIMIT 50

OBSERVE

97.8% of slow spans run on inventory-shard-19 after deploy 8c41fd.

ANSWER

Probable cause found in 3 queries. Recommend draining shard-19 and reverting inventory deploy 8c41fd.

[ DIFFERENT CONSTRAINTS ]

Stop paying the
dashboard tax.

Traditional observability optimizes for prebuilt views and human browsing. TraceStore retains raw context forever and keeps the bill to two predictable inputs: storage and query.

CONSTRAINTTRACESTORE / AGENT-NATIVEDASHBOARD-NATIVE

Primary user+Agents−Humans

Retention+Infinite by design−Priced expiry tiers

Cost model+Storage + query−Hosts + seats + signals + indexes

High cardinality+Context to preserve−A tax to control

Interface+Composable SQL−Dashboards + proprietary UI

[ PRIVATE ALPHA / DESIGN PARTNERS ]

Your telemetry has a
new primary user.

We’re working with a small group of AI infrastructure teams. If your agents need complete production history without an unpredictable observability bill, we should talk.