Experience, Harbor Trajectories, and OpenInference

Experience is Sikaru’s agent trajectory abstraction. It captures the episode-level data that continual learning and agent eval systems need: messages, tool calls, tool results, telemetry events, outcomes, corrections, failures, and rewards. Sikaru Experience is compatible with the broader agent trajectory ecosystem:

Harbor ATIF trajectories can be uploaded to Sikaru.
Sikaru trajectories can export to Harbor ATIF v1.7.
OpenInference spans can be uploaded directly.
OTLP resource spans can be normalized into the same trace stream.
Native Sikaru Experience payloads, Harbor ATIF, and OpenInference-compatible traces all feed the same issue mining, eval seed, and improvement evidence loop.

Experience and traces write to POST /v1/trace-streams.

Why Experience matters

Managed agents improve when Sikaru can connect what happened to whether it was good. A single Experience can carry:

Signal	Why it matters
Messages	The user request and assistant behavior.
Tool calls and results	The product capabilities the agent used and what came back.
Feedback and corrections	Human-provided guidance about the better behavior.
Failures	Known misses, exceptions, timeouts, rejected answers, or policy issues.
Outcomes and rewards	Task success, acceptance, escalation quality, or other product-defined measures.
Metadata	Project, account, user, conversation, environment, source, and trace identifiers.

This is the evidence Sikaru uses to find repeated failures, select or create eval cases, and propose reviewed behavior updates.

Record a product-run experience

Use Experience when your product already executed the agent behavior itself.

Python
TypeScript

experience = sikaru.experience(
    "support",
    user="user_123",
    conversation="conv_123",
    session="session_123",
)

experience.user("Find renewal risk.")
call = experience.tool_call("crm.search", {"account_id": "acct_123"})
experience.tool_result(call, {"risk": "legal review"})
experience.assistant("Renewal risk is legal review.")
experience.signal("approval.completed", {"approved": True})
experience.outcome("operator_acceptance", 1.0, feedback="accepted")

result = experience.commit(dataset="live")

const experience = sikaru.experience("support", {
  userId: "user_123",
  conversationId: "conv_123",
  sessionId: "session_123",
});

experience.user("Find renewal risk.");
const call = experience.toolCall("crm.search", { account_id: "acct_123" });
experience.toolResult(call, { risk: "legal review" });
experience.assistant("Renewal risk is legal review.");
experience.signal("approval.completed", { approved: true });
experience.outcome("operator_acceptance", 1, { feedback: "accepted" });

const result = await experience.commit({ dataset: "live" });

Add corrections and failures

Corrections and failures make the continual learning loop actionable. Use corrections when a user, operator, evaluator, or policy reviewer identifies the better behavior. Use failures when a run produced a known miss, exception, policy issue, timeout, or rejected answer.

Python
TypeScript

experience.correction("Use legal-review wording before recommending renewal terms.")
experience.failure("retry_needed", message="first answer was vague")

experience.correction("Use legal-review wording before recommending renewal terms.");
experience.failure("retry_needed", { message: "first answer was vague" });

Build a Sikaru trajectory from existing messages

Use this path when another agent framework already produced messages and tool calls.

from sikaru_sdk import build_trajectory_from_messages, messages_from_openai_chat

trajectory = build_trajectory_from_messages(
    messages_from_openai_chat(messages),
    conversation_id="conv_123",
    trace_id="trace_123",
    data_source="openai-agent",
)

sikaru.traces.upload_trajectory(trajectory, dataset="prod")

Upload OpenInference-compatible traces

OpenInference spans are useful for migrations, low-level instrumentation, and existing tracing pipelines.

Python
TypeScript

sikaru.traces.upload(openinference_spans, dataset="prod")
sikaru.traces.upload_file("./openinference-export.jsonl", dataset="prod")

await sikaru.traces.upload(openinferenceSpans, {
  dataset: "prod",
});

Use deterministic idempotency keys for retried single-batch uploads.

Python
TypeScript

sikaru.traces.upload(
    openinference_spans,
    dataset="prod",
    idempotency_key="trace-batch-2026-06-14-001",
)

await sikaru.traces.upload(openinferenceSpans, {
  dataset: "prod",
  idempotencyKey: "trace-batch-2026-06-14-001",
});

Upload OTLP and Harbor ATIF

The Python SDK has helpers for direct OpenTelemetry OTLP resource spans and Harbor ATIF trajectories.

sikaru.traces.upload_otlp_resource_spans(
    otlp_payload["resourceSpans"],
    dataset="prod",
    metadata={"source": "otel-exporter", "account_id": "acme"},
)

sikaru.traces.upload_harbor_trajectory(
    atif_trajectory,
    dataset="harbor-evals",
    metadata={"source": "harbor", "account_id": "acme"},
)

Export a Sikaru trajectory to Harbor ATIF v1.7 when you want to run or share agent trajectory evals outside Sikaru.

atif = trajectory.to_harbor_atif(
    agent_name="support",
    agent_version="2026.7.3",
)

Metadata for evals and mining

Attach stable metadata that helps Sikaru group, search, and explain issues.

Metadata	Typical value
`project_id`	Added by the SDK from the configured project.
`account_id`	Customer, tenant, or workspace account.
`environment`	`production`, `staging`, or another deployment label.
`source`	Agent framework, service, or SDK source name.
`user_id`	Product user ID.
`conversation_id`	Conversation, ticket, thread, or workflow ID.

Treat outcomes, corrections, and failures as product facts. Sikaru uses them as evidence for issue mining, eval seed selection, regression checks, and reviewed improvements.

​Why Experience matters

​Record a product-run experience

​Add corrections and failures

​Build a Sikaru trajectory from existing messages

​Upload OpenInference-compatible traces

​Upload OTLP and Harbor ATIF

​Metadata for evals and mining

Why Experience matters

Record a product-run experience

Add corrections and failures

Build a Sikaru trajectory from existing messages

Upload OpenInference-compatible traces

Upload OTLP and Harbor ATIF

Metadata for evals and mining