Skip to content
Open agent telemetry for Confluent Cloud
Paper ComputePaper Compute
×Confluent

Traces that fuel Confluent Intelligence.

One agent produces a trace. A production fleet produces a continuous stream of decisions, tool calls, outcomes, and failures. Paper Compute captures it. Confluent makes it durable, queryable, and actionable.

Open source captureNo SDKs. No code changes.
Fleet-scale streamSessions, spans, stats, and anomalies.
Your Confluent CloudYour topics, ACLs, and retention.
01

The Problem

AI agents continuously make LLM calls, invoke tools, access data, and take action. As customers move from a pilot to a fleet, that activity becomes a large, fast-moving behavioral dataset. Most of it disappears after each run, leaving operations and compliance without a fleet-wide record. The fix shouldn't be adopting another dashboard.

Infrastructure shows
Resource metrics
CPU, memory, latency
Agents generate
Behavioral data
Decisions, actions, outcomes
Confluent unlocks
Fleet intelligence
Patterns across every agent
02

Open Core

tapes is open source infrastructure for agent telemetry. Cassettes are the components you attach to it—the Confluent cassette is a metered sidecar that streams derived telemetry into your Confluent Cloud. Deploy it yourself, or let Paper operate it for you.

tapesOpen Source

Zero-instrumentation capture at the network layer. Records every LLM call, tool invocation, and execution path, then derives sessions, spans, and stats from the raw record. No SDKs. No code changes. One command.

$ tapes start
Proxy listening on :4020
Recording locally, deriving sessions...
14 sessions · 2,847 requests · 1.2M tokens
Confluent cassetteMetered

The streaming export. A sidecar next to tapes that produces derived events—sessions, spans, stats, anomalies—to topics in your Confluent Cloud. One cassette, one config, no bespoke connectors. Raw session content never leaves your boundary.

$ tapes cassette add confluent
bootstrap  pkc-xxxxx.confluent.cloud:9092
topics     agent.sessions.v1 · agent.telemetry.alerts
policy     derived events only
✓ streaming — raw sessions stay local
Prefer not to run it?

Managed Paper operates the capture gateway and cassettes for you—same open schemas, same topics in your cluster, zero ops on your side.

03

Where Confluent Fits

As agent counts and run volumes grow, Paper Compute turns activity from every agent into one consistent event stream. Confluent Cloud—topics, Flink, Tableflow, and existing dashboards—is where that aggregate becomes durable, analyzable, and valuable.

AI Agentsany frameworkYOUR INFRASTRUCTUREGatewaycapture · your cloudtapes coreraw → derivedopen sourceConfluent cassettemetered sidecarnever raw sessionsevents →← contextCONFLUENT CLOUDKafka topicsyour clusterFlink SQLanomaliesTableflowaudit tablesRead surfaceyour console
01Kafka TopicsConfluent Cloud

High-volume agent telemetry lands as structured events—sessions, spans, stats, and anomalies—on topics in your cluster. Instead of isolated traces, customers get a durable behavioral stream across every agent, governed by their retention, ACLs, and schema registry.

02Apache FlinkConfluent Cloud

Anomaly detection purpose-built for agent failure modes—stuck loops, token spikes, behavioral drift—runs across the live fleet, not one trace at a time. The Flink SQL is readable, editable, and runs in the customer's Confluent environment. Not ours.

03The CassetteMetered

One export cassette, configured for Confluent. It sits next to tapes, reads the derived stream, and produces to your topics. The same cassette design targets other sinks with a config change—never a bespoke connector per destination.

04Confluent IntelligenceContext Loop

Confluent's Real-Time Context Engine feeds live context into agents. tapes captures what agents do with that context, and the cassette streams it back. Context in, telemetry out—a closed observability loop built on Confluent primitives.

The data boundary
Flows to Confluent

Fleet-wide projections: sessions, spans, stats. Judgment outputs: anomalies, outcomes. Continuous, structured, schema-registered, and queryable at scale.

Stays in your boundary

Raw session content: prompts, responses, tool payloads. Recorded by tapes, kept wherever tapes runs. Projections can be rederived from it at any time.

Compliance reads the derived stream in Confluent. The raw record never has to leave.

04

Real-Time Anomaly Detection

Purpose-built for agent failure modes, evaluated continuously across the fleet: Flink SQL finds stuck loops, token spikes, and drift in the combined stream, then writes alerts to a topic in the customer's cluster.

flink sql · stuck-loop job● running
-- same tool call repeating inside a window
INSERT INTO agent_telemetry_alerts
SELECT session_id, 'STUCK_LOOP', COUNT(*)
FROM TABLE(TUMBLE(TABLE agent_events,
  DESCRIPTOR(event_time), INTERVAL '60' SECOND))
WHERE event_type = 'tool_call'
GROUP BY session_id, tool_hash, window_start
HAVING COUNT(*) >= 5;
agent.telemetry.alerts · consumerlive
12:47:03  STUCK_LOOP     deploy-agent-prod  5x same call / 60s
12:49:41  TOKEN_SPIKE    code-review-agent  2.3x rolling avg
12:51:22  NO_PROGRESS    sweeper-agent      100+ events, 5 states
→ advice inbox: patch applied to live run
alerts also land in observations the agent loads at start
Stuck Loops
Same tool call 3+ times
Token Spikes
2x rolling average
Behavioral Drift
Output divergence
Prompt Injection
Input pattern analysis

Alerts don't stop at a dashboard. A consumer can write them back to a running agent as advice—the loop below is live code, not a diagram.

05

Proof: A Self-Healing Agent on Kafka

This architecture isn't a slide. An autonomous agent plays Pokémon Red headlessly inside a stereOS VM, and every battle, map change, and stuck loop streams through the exact pipeline on this page—while Paper records what the agent was thinking on the side.

pcc-labs/pokemon-kafka · pipeline● live
agent → game events (pokemon.game.v1)
  └→ Kafka (agent.game.events)
     ├→ Flink SQL → agent.telemetry.alerts
     │    └→ advice inbox → agent applies mid-run
     └→ dlt → warehouse (DuckDB / Snowflake)
Paper records LLM sessions separately —
what it thought (Paper) vs. what it did (Kafka)
flink sql · anomaly jobs7 running
GAME_STUCK_LOOP60s · 5+ stuck events on one map
BATTLE_WIPE5min · player HP hits 0
POSITION_DEADLOCK2min · 50+ events at one tile
NO_PROGRESS5min · 100+ events, ≤5 tiles
BATTLE_LOSS_STREAK10min · 2+ lost battles

The loop closes. Flink alerts become advice a live run polls between turns—parameter patches hot-applied mid-game, no operator involved. When tuning is exhausted, the evidence escalates to an engine that proposes a code change, runs the gates, and opens a PR. Telemetry out, healing in: the closed loop this page describes, running on a real workload.

pcc-labs/pokemon-kafka
06

Integration Roadmap

Confluent Connector Hub

List the Confluent cassette on the Hub so any Confluent customer installs it like a connector. The open topic schemas are the contract.

Tableflow

Turn the growing telemetry stream into queryable tables so operations, compliance, and security can analyze behavior across the fleet without engineering involvement.

Real-Time Context Engine

Feed fleet-wide anomaly data back into the context engine so agents self-correct based on failures observed across every run—the advice-inbox pattern on Confluent primitives.

07

The Opportunity

One agent creates a trace. A fleet creates a continuously growingbehavioral dataset.

Every new agent workload creates another stream of decisions, tool calls, outcomes, and failures. Confluent customers can turn that volume into two answers their teams need:

Paper Compute supplies
What did every agent actually do?
Confluent unlocks
What patterns are emerging across the fleet?

The value compounds with scale. Every run enriches a durable dataset for real-time detection, historical analysis, benchmarking, capacity planning, and audit. It lands where customer teams already work—no new dashboard to roll out and no new vendor UI to learn.

Together, Confluent and Paper Compute give enterprises full-stack observability for agent fleets. Regulated industries—financial services, healthcare, defense—that already require audit trails for data pipelines will require the same for the agents operating on that data. Paper Compute creates the behavioral stream; Confluent makes it operationally useful; the raw record never leaves the customer's boundary.

Financial Services
Regulatory audit trails
Healthcare
HIPAA compliance
Defense
Decision accountability

Bring agent telemetry into the Confluent conversation.

When a customer is moving from one agent to a production fleet, every run becomes valuable streaming data. Paper Compute captures it and Confluent turns the volume into real-time insight. Bring us into the deal for a demo, architecture review, or joint customer call.

Confluent was acquired by IBM in 2025. For enterprise buyers, this means Confluent Cloud is now backed by IBM’s global support, compliance certifications, and procurement infrastructure. Paper Compute’s integration targets Confluent Cloud APIs and is unaffected by the change in ownership.