Product · AgentSight v1.0.25

Start with the fleet or machine. Drill into one agent session when you need the details.

AgentSight v1.0.25gives signed-in users an All machines organization view, while a Direct or local Node keeps its own machine overview. Select a Node and then a session to open Conversation, Process Tree & AI Prompts, or Analysis. Portable agent-native history works on Windows, macOS, and Linux; Linux can add eBPF-backed system capture.

Open AgentSight Read the documentation

All machines

The signed-in organization view aggregates reachable Nodes in the browser, not in a cloud telemetry database.

The current hosted app queries each reachable Node through Direct or Relay, then combines bounded machine overviews in browser memory. Filters and a machine selector let you move from the organization view to one Node without copying Node snapshots into Controller storage.

Machines and sessions

See reachable machines plus active and stopped agent sessions before choosing one Node.

Observed usage

Compare reported Tokens, source-reported subscription windows, and Agent Plans without deriving provider quota from captured tokens.

Machine resources

Keep CPU, RSS, process counts, and session state beside the agent work they describe.

One Node

The machine overview stays bounded; transcript detail waits behind session selection.

A Node overview answers what is running, what stopped recently, how observed Tokens and source-reported capacity look, which Agent Plans are active, and what CPU/RSS/process shape the machine has. The frequent refresh path does not need to load every complete transcript.

01
Choose a machine
Use All machines, Direct, or a local Node depending on how you connected.
02
Choose a session
Select the task that actually needs debugging, review, or comparison.
03
Load detail
Fetch conversation and session-scoped runtime data only for that session.

Session workspace

Three views share the same selected-session boundary.

v1.0.18 replaced separate Timeline and Detailed Events tabs with one Analysis workspace. Raw event detail is still available by selecting an event in the interactive timeline, while summary metrics stay tied to the same session rather than becoming another global dashboard.

Session view

Conversation

Read the selected session’s user and assistant messages plus tool activity.

Session view

Process Tree & AI Prompts

Connect the chosen session to its process family and captured model context.

Session view

Analysis

Summarize duration, LLM turns, failures, models, tokens, tools, files, network targets, processes, resources, and an interactive event timeline.

Observed Tokens versus subscription capacity

AgentSight keeps measured usage separate from provider-reported limits.

The machine and fleet views can show observed token use beside source-reported subscription metadata when the local agent source provides it. For Codex, AgentSight allowlists fields such as plan type and rate-limit windows from local source-native metadata. It does not infer provider quota from captured token totals, and Controller does not persist these Node-provided usage details.

Keep the sources separate

Observed usage

What AgentSight saw in the session.

Reported capacity

What the local agent source reported about its own limit window.

Portable sessions versus system capture

Cross-platform session management does not mean cross-platform eBPF tracing.

Agent-native session parsers support top, bind, vis, and report workflows on Windows, macOS, and Linux without eBPF. Linux adds the independent record and eBPF-backed debug path when an investigation needs process, file, resource, or TLS detail beyond the agent-native transcript.

Native Windows support does not add an ETW/eBPF capture backend or claim Linux tracing parity. The v1.0.25 GitHub release publishes Linux x86_64 and aarch64 binaries; Windows builds are exercised by CI and can be built from source or obtained as CI artifacts rather than a release asset.

Choose by question

Session history

Start with the portable agent-native source.

Host effects

On Linux, add system capture when the transcript cannot answer the question.

Live Node versus saved capture

A saved database is useful history, but it is not a live machine.

A live Node can provide current machine state, fresh session detail, and supported session messaging. A server opened over a saved SQLite capture remains read-only and cannot expose live state that was not present in the saved artifact.

Unassigned saved, demo, or live runtime data remains useful through recorded-capture views, but negative claims should stay bounded by the source actually opened.

Record a bounded run Review data handling

Distributed by transport, local-first by data

Direct and Controller relay are transports to Nodes, not new session databases.

Direct requests travel browser to Node. Controller relay remains an optional managed path. The hosted app can aggregate bounded Node overviews in browser memory, while Controller keeps identity, organizations, machine directory, connectivity, and access policy. Architecture owns the exact data and trust boundary.

01
Local
Use AgentSight on one machine without a hosted account.
02
Direct
Reach a browser-visible Node through its explicit endpoint.
03
Controller-managed
Add identity, organizations, discovery, and optional relay when needed.

Hosted sign-in status

Supported sign-in choices stay visible even when a provider is not configured.

v1.0.21 keeps GitHub and Google entries visible when hosted sign-in is available. If a provider is not configured on Controller, the UI disables it and explains the configuration state instead of silently removing it. A Controller provider-status failure is surfaced as an error. This does not claim that every deployment has both providers configured.

Open the hosted app Read the architecture

Product map

Go deeper without duplicating the same answer.

Product owns the end-to-end workflow. Architecture owns the data/control boundary, Integrations own runtime compatibility, Use cases own engineering decisions, and Compare pages own adjacent observability boundaries.

Use cases

Start from the engineering question: slow runs, generated changes, closed-source agents, or extension audits.

Architecture

Understand the Node, Direct, Controller, organization, browser aggregation, and capability boundaries.

Integrations

Choose the capture path that matches the agent and runtime you already use.

Compare

Compare AgentSight with application traces, OpenTelemetry, gateways, and hosted LLM tools.

Primary sources

This workflow is pinned to the released v1.0.25 implementation.

Reviewed on 16 August 2026 against AgentSight v1.0.25 at 0080545f7c6b110ec2d4a4af5100b58f514c84d5. The fleet, usage, and session-analysis surfaces changed materially after v1.0.17, so use the exact release and source links below when a deployment or audit depends on a specific version.