Coding Horizon

Alibaba Just Made AI Agents Much Harder To Hide

Sources for every figure, version, requirement and capability the finished picture puts on screen. Chased to a primary source and cited.

What AgentSight is

Alibaba documents AgentSight as an agent observability component inside ANOLISA, its system analysis toolkit for Anolis OS. The documentation opens:

“AgentSight is a zero-instrumentation AI Agent observability tool based on eBPF. It captures LLM API calls, Token consumption, and process behavior at the kernel level without modifying Agent code.”

Source: alibaba/anolisa, docs/user-guide/en/agent-observability/agentsight.md https://github.com/alibaba/anolisa/blob/main/docs/user-guide/en/agent-observability/agentsight.md

The upstream project of the same name is an open source research tool from the eunomia-bpf group, MIT licensed, first published July 2025, with an accompanying paper.

Capabilities on screen

The ANOLISA documentation lists these capabilities in a table, and the video’s shots follow it item for item.

On screen Documented as
token consumption analysis “Multi-dimensional Token accounting by agent, task, and model”
behaviour audit “Complete tracing of LLM calls and process execution”
dashboard “Web UI for real-time Token trends, Agent health, and session traces”
agent discovery “Automatic detection of running AI Agent processes”
interruption detection “Detection of LLM errors, SSE truncation, context overflow, and crashes”
external log export “Supports exporting structured events to external log services”

Source: as above.

The five interruption types

The documentation gives the type names and their default severities, which the shot renders exactly:

Type Condition Default severity
llm_error HTTP status >= 400, or an SSE body containing an error high
sse_truncated SSE stream ended without finish_reason=stop high
context_overflow context length exceeded high
agent_crash agent process disappeared mid session critical
token_limit finish_reason=length with output near max medium

Source: as above.

Requirements, which the video states as the catch

On screen Documented as
Linux only “OS: Linux”
kernel 5.8 or newer, with BTF “Kernel: >= 5.8 (BTF support required)”
root or CAP_BPF “Privileges: root or CAP_BPF (for eBPF probes)”
a system service sudo systemctl enable --now agentsight.service is the recommended deployment
data under a system log path “stores data under /var/log/sysak/.agentsight”, root owned, private umask
dashboard port http://localhost:7396

Source: as above.

The upstream repository states a lower kernel floor for its own binary (“Linux kernel: 4.1+ with eBPF support (5.0+ recommended)”), because it ships a different set of commands. The video quotes the Alibaba documentation, which is the packaging being discussed, and that documentation says 5.8 with BTF.

The macOS limitation

“On macOS, AgentSight provides two commands: trace (trajectory collector that scans local JSONL session files, no eBPF) and serve (Dashboard viewer). All other eBPF-dependent commands are Linux-only.”

Source: as above. The shot shows trace and serve lit and the remaining documented commands (token, audit, discover, interruption) dark.

What an event actually carries

The video renders an audit event with pid, comm, command, argv, model, input_tokens and output_tokens, and states that command line arguments can be exposed. The upstream snapshot schema documents those fields, and flags argv explicitly:

argv — array of strings — “Yes; may include user data, paths, tokens, or secrets.”

It also documents command (“may contain the executable or command text”), comm, pid, ppid, root_pid, and a model identifier on token records.

Source: eunomia-bpf/agentsight, docs/snapshot-schema.md https://github.com/eunomia-bpf/agentsight/blob/master/docs/snapshot-schema.md

The comparison the video draws

The upstream README makes the same comparison the script makes, naming the tools:

“Application-level tools such as LangSmith, Langfuse, and Phoenix are great for traces, prompts, tokens, evals, and latency when you own the application code… AgentSight focuses on the layer those tools often miss: what the agent actually does at the system boundary.”

It also states the mechanism: plaintext is captured at SSL/TLS call boundaries rather than through a proxy, and process execution and file access are correlated with it.

Source: https://github.com/eunomia-bpf/agentsight

Overhead

The paper reports the boundary tracing approach costs “less than 3% performance overhead”. The video does not put this number on screen, because the narration does not claim it.

Source: arXiv:2508.02736.

Not checked