Observe runtime evidence
Connect prompts, model and tool activity to processes, files, network effects, resources, and the rest of the real run.
New AgentSight v1.0.15 is available →
Agent-native observability and evolution
Observe and evolve your AI agents with AI agents.
AgentSight turns agent trajectories and system-level runtime evidence into a closed improvement loop. AI agents can diagnose repeated failures, generate skill improvements, compare baseline and candidate behavior, and promote or roll back changes.
Runtime evidenceTrajectory diagnosisSkill evolutionEval + rollback
AgentSight live sessionsHosted app + Direct Node
Works with the agents and runtimes you already use
Claude CodeCodexGemini CLIOpenCodeOpenClawPythonNode.jsContainers
v1.0.15Current release
MITOpen-source license
arXiv + ACMPublished system research
OTel GenAIModel-call export
Local firstSQLite session artifacts
Observe → diagnose → evolve → evaluate
Close the agent improvement loop.
AgentSight connects what agents intended to what actually happened, finds repeated failure patterns, turns them into reusable skill changes, and validates candidate behavior before promotion.
Diagnose repeated failures
Mine agent trajectories for recurring mistakes, correction loops, friction, and the evidence behind them.
Evolve agent skills
Turn repeated failures into small, versioned skill changes and reusable improvements instead of one-off retries.
Evaluate, promote, or roll back
Compare baseline and candidate behavior on held-out tasks, guard regressions, promote improvements, and roll back bad changes.
Explore the product
Move from a busy run to a clear explanation.
Start with the Overview dashboard, then use the timeline, process tree, resource views, repository replay, and flamegraphs to inspect the part of the run that matters.
01
Overview dashboard
Start with the whole run.
Review duration, tokens, model calls, processes, files, endpoints, activity over time, resource shape, and friction signals before drilling into detailed views.
Open this view
02
Timeline
Follow the run in order.
Line up model calls, tool activity, processes, file operations, and network events on one timeline.
Open this view
03
Process tree
See which agent launched each command.
Inspect child processes and file activity under the agent session that caused them.
Open this view
04
Resource metrics
Find expensive and stalled phases.
Compare CPU and memory behavior with the model and tool activity happening at the same time.
Open this view
Repository replay
Watch how an agent changed a codebase.
Agent Nebula replays file reads, writes, creates, renames, and deletes across a Git worktree.
Agent Flamegraph
Find where tokens and time went.
Aggregate real local agent sessions by project, agent, prompt category, model, and token type.
Product screenshots and examples are pinned to AgentSight commit ba14044491d2fdb52e8b9d0f3e9a94c5d3a12dd1.
After the run
Query or open the session from several angles.
A recorded run is a reusable local artifact. Use focused report commands, local or Direct Node web views, repository replay, and profiling views without turning the product website into duplicate CLI documentation.
agentsight report audit --jsonAudit process, file, and API activity
Inspect process spawns, file opens, and captured API activity from a saved run instead of relying on the final agent answer alone.
Review data and security →agentsight report tokenProfile token use across sessions
Summarize token usage from the latest AgentSight database or supported local agent sessions, then move to Agent Flamegraphs for deeper aggregation.
Explore token profiling →agentsight bindOpen local data in the hosted app
Start an unprivileged loopback Direct Node with a process-lifetime access key. Detailed session data stays on the Node while the hosted app reads it directly.
Open the AgentSight app →agentsight report serveReopen a recorded session in the local web UI
Use the same Overview, timeline, process-tree, log, and metrics views on saved SQLite sessions after the original command has finished.
Open the retained recorded demo →agentsight visReplay repository changes
Turn local Claude, Codex, and Gemini session history into an Agent Nebula replay of reads, writes, creates, renames, and deletes in a Git worktree.
See the review workflow →Built for real engineering work
Debug slow runs, review changes, and inspect agent tools.
Use AgentSight when the final answer or code diff is not enough to explain how the agent reached it or what it changed along the way.
Performance use case
Profile slow or expensive AI agent runs
Find where an AI agent run spent time and tokens by connecting model calls to commands, files, waits, and process activity.
Read the pageCode review use case
Review AI-generated pull requests with execution history
Review an AI-generated change with the commands, tests, retries, file effects, and network activity that produced it.
Read the pageCompatibility use case
Trace closed-source AI agent CLIs
Observe local AI agent commands and child processes without requiring an SDK, source change, proxy, or vendor-specific callback.
Read the pageExtension audit use case
Audit MCP servers, skills, and agent plugins
Run an agent extension under system-level tracing to inspect its subprocesses, file changes, remote calls, and resource cost.
Read the pageStandards comparison
AgentSight and OpenTelemetry
Understand how AgentSight system profiles relate to OpenTelemetry instrumentation, traces, semantic conventions, and backends.
Read the pageClaude Code integration
AgentSight for Claude Code
Record Claude Code from the system boundary and connect model activity to commands, files, processes, network calls, and resources.
Read the pageResearch and operating boundaries
Technical claims should be inspectable.
AgentSight is an open-source systems project with a published paper, versioned source, public product artifacts, and explicit data-handling limits. The website should expose those facts directly.
Native agent telemetry vs. system observation: where does the boundary actually sit?
We reviewed current Claude Code telemetry, Gemini CLI at an exact commit, Codex OpenTelemetry source, OpenTelemetry GenAI conventions, MCP, and AgentSight. Native telemetry is already rich; independent system observation becomes most useful when execution crosses into descendant processes and low-level effects.
Read the boundary map →AgentSight: System-Level Observability for AI Agents Using eBPF
The AgentSight paper describes boundary tracing, correlation across model and system activity, the system design, and the evaluation behind the project.
Read the paper on arXiv →Recorded sessions are useful because they are detailed and should be treated as sensitive.
Session databases and exports can contain prompts, responses, paths, headers, and network targets. Keep raw artifacts local or handle them with the same care as other development telemetry.
Read Security and data handling →Open source and local first
Run it around the command you already use.
AgentSight records locally, works with closed-source CLIs, and exports captured model calls as OpenTelemetry GenAI spans when you want to connect them to an existing telemetry stack.
- ✓ Existing CLI and terminal workflow
- ✓ Local SQLite sessions and saved-session web UI
- ✓ Direct Node access without uploading detailed session contents
- ✓ eBPF process and file monitoring
- ✓ TLS tracing without a model proxy
- ✓ Evidence-gated skill evolution with evaluation and rollback
01
agentsight top02
sudo agentsight record -- claude03
agentsight report audit --json04
agentsight report token05
agentsight bindTry AgentSight
Observe a run. Evolve the next one.
Open the hosted app, connect a Direct Node, or explore the retained recorded session. AgentSight keeps detailed runtime evidence available for diagnosis, skill evolution, evaluation, and rollback.