Open-source platform purpose-built for AI agent debugging. Trace multi-step reasoning, attribute costs per task, score quality at the outcome level.
from agentscope import AgentScope
scope = AgentScope.init(
endpoint="http://localhost:3001",
project_id="my-project",
)
@scope.wrap_agent("research-run", agent_name="researcher")
async def research(task: str):
with scope.trace_step("planning"):
plan = await llm.chat(...)
with scope.trace_tool("search"):
data = await search(query=plan.q)
return dataPurpose-built tools for understanding, measuring, and improving AI agent behavior in production.
Multi-step traces with tool calls, sub-agent delegation, and decision trees. Not just LLM call logging.
Know exactly what each agent costs per task, per tool, per model. Set budgets with auto-pause guardrails.
Score agent outputs at the outcome level. Automated evaluators, human review, and composite scores.
Convert production agent failures into regression test cases with one click. Close the non-determinism loop.
Step-by-step visual replay of agent decision trees. Compare successful vs failed runs.
Built on OTel standards. Works with any collector, any backend. No vendor lock-in.
AgentScope is designed to support AI governance programs with local processing options, sovereignty-aware data handling, and audit-oriented operational visibility.
Self-hosted deployment keeps trace ingestion, storage, and analysis inside your infrastructure boundary.
Tenant and project boundaries let teams align observability data handling with internal residency policies.
Trace timelines, cost records, and quality signals help teams build investigation and governance evidence trails.
Compliance posture depends on deployment configuration and policy setup.
See how AgentScope compares to existing tools
| Feature | AgentScope | Others |
|---|---|---|
| Agent-native tracing | Partial | |
| Cost per agent/task | Manual | |
| Failure-to-test | ||
| Decision replay | ||
| Budget guardrails | ||
| MIT License | Varies | |
| Self-hosted | Varies |
Three steps to full agent observability.
$ docker compose up -dfrom agentscope import AgentScope
AgentScope.init(endpoint="http://localhost:3001")Open localhost:3000 to see your traces, costs, and quality scores in real-time.
Start free, scale when you need to.
Deploy in minutes. Open source, self-hosted, no vendor lock-in.