Kickstarting the Data Flywheel: Four Ways to Connect Agents to AgentLoop

Data ingestion is the starting point of the data flywheel; without high‑quality runtime data, observation, evaluation, and experiments are impossible. Connecting an Agent is not trivial because agents can be generic, framework‑based, custom‑coded, or require zero‑code changes.

Design of AgentLoop data ingestion

AgentLoop builds its data‑ingress capability on the OpenTelemetry (OTel) standard protocol and operates as a probe. The probe captures every model call and tool call inside the Agent, links them with a Trace ID, and constructs a complete trace topology that represents the execution chain.

Why OTel + probe?

OTel is the de‑facto standard in observability, guaranteeing that data can be consumed by any OTel‑compatible ecosystem. The probe, refined over more than a decade, gathers data from all corners of the program without modifying business logic, providing the core competitive advantage of AgentLoop’s ingestion layer.

Four integration methods

1. Generic Agent – one‑click onboarding : Specialized instrumentation for common out‑of‑the‑box agents; no development work is required.

2. Agent framework – SDK integration : For agents built on frameworks such as AgentScope, the SDK automatically instruments model and tool calls, making data collection invisible to developers.

3. Custom high‑code – annotation : Developers add annotation‑style markers at the exact points they want to observe; with AI‑coding the extra lines are minimal.

4. eBPF – zero‑intrusion : A kernel‑level monitor that requires no code changes, ideal for legacy, closed‑source, or exploratory scenarios.

Key concepts

Service Name : Identifies the originating application within a shared AgentSpace, enabling scoped observation and evaluation.

License Key : Serves as an authentication token; without it, data cannot be uploaded.

Practical demo: Connecting a Claude Code customer‑service Agent

The demo uses a customer‑service Agent built with Claude Code and the Claude Agent SDK. Two entry points exist: (1) the probe method and (2) a Claude Code webhook. The demonstration follows the probe method combined with the LoongSuite Pilot side‑car process.

Step 1 – Obtain License Key and run the install command

A License Key is fetched from the console, then a single install command is executed. The command first installs a dependency and then deploys the LoongSuite Pilot collector; the entire setup requires only this one command.

Step 2 – Understand the collection mechanism

LoongSuite Pilot runs as a side‑car (bypass) process. Claude Code’s webhook forwards data to Pilot, which then reports it to the cloud. Because Pilot is out of the Agent’s execution path, collection overhead and failures never affect the Agent’s normal operation.

Step 3 – Copy configuration and restart the Agent

After installation, a settings.json file containing the hooks is generated. This file must be copied into the Agent’s installation directory and the Agent restarted so that the hooks become effective. Forgetting either step prevents data upload.

Step 4 – Test and verify on the observation page

A test request is sent to the Agent; the Agent responds normally. The observation page is then opened, confirming that data has been reported. The trace view shows the full execution chain, including both tool calls and the large‑model processing step, proving that the probe captured the entire workflow.

The complete ingestion flow—from a single install command to a verified trace—takes less than ten minutes. With the data loop closed, the platform now has a continuous supply of high‑quality data for the next stage of the flywheel.