Import ATIF Trajectories

Upload agent trajectories in ATIF (Agent Trajectory Interchange Format) to Phoenix and visualize them as traces.

ATIF (Agent Trajectory Interchange Format) is an open schema for recording agent execution traces. Agent frameworks like Claude Code, OpenHands, Gemini CLI, and Codex export ATIF via Harbor, a benchmark harness for evaluating coding agents.

The phoenix-client package includes a utility that converts ATIF trajectory JSON into OpenTelemetry-compatible span trees and uploads them to Phoenix, so you can visualize and evaluate agent runs using Phoenix’s tracing UI.

Quick start

import json
from phoenix.client import Client
from phoenix.client.helpers.atif import upload_atif_trajectories_as_spans

# Load one or more ATIF trajectory dicts
with open("trajectory.json") as f:
    trajectory = json.load(f)

client = Client()
result = upload_atif_trajectories_as_spans(
    client, [trajectory], project_name="my-agent-eval"
)
print(result)  # {"total_received": 5, "total_queued": 5}

If you’re using Phoenix Cloud, set PHOENIX_CLIENT_HEADERS and PHOENIX_COLLECTOR_ENDPOINT before creating the client. See Connect to Phoenix.

Trace hierarchy

ATIF stores agent steps as a flat list. The converter builds a hierarchical span tree that matches what real-time instrumentors produce:

Single-turn conversations (one user message):

AGENT (root - input=user message, output=final agent reply)
  LLM
  TOOL
  LLM

LLM and TOOL spans are siblings under the root AGENT span. The agent runtime executes tools, not the LLM, so tool spans are peers of LLM spans rather than children.

Multi-turn conversations (multiple user messages):

AGENT (root - input=first user message, output=final agent reply)
  AGENT turn_1 (input=user msg 1, output=agent reply 1)
    LLM
    TOOL
  AGENT turn_2 (input=user msg 2, output=agent reply 2)
    LLM

Each follow-up user message starts a new turn, represented as a nested AGENT span under the root.

Batch uploads and subagent linking

When an agent delegates work to a subagent, the ATIF trajectories reference each other via subagent_trajectory_ref. Upload the parent and child trajectories together in one call and the converter automatically nests the child’s spans under the parent’s tool span:

with open("parent_trajectory.json") as f:
    parent = json.load(f)
with open("child_trajectory.json") as f:
    child = json.load(f)

# Upload together so cross-references resolve
upload_atif_trajectories_as_spans(
    client, [parent, child], project_name="my-agent-eval"
)

The resulting trace looks like:

AGENT (parent)
  LLM
  TOOL (delegate_task)
    AGENT (child agent)
      LLM
      TOOL

Continuation merging

When an agent’s context window fills up, Harbor splits the session across multiple trajectory files. The continuation file gets a session_id ending in -cont-N. The converter automatically detects these and merges them into a single trace, so the full agent session is visible as one trace in Phoenix.

Continuation root spans are annotated with metadata.is_continuation = True.

Attribute mapping

The converter maps ATIF fields to standard OpenInference attributes:

ATIF field	OpenInference attribute
metrics.prompt_tokens	llm.token_count.prompt
metrics.completion_tokens	llm.token_count.completion
metrics.cached_tokens	llm.token_count.prompt_details.cache_read
metrics.cost_usd	llm.cost.total
agent.model_name / step model_name	llm.model_name
agent.tool_definitions	llm.tools.{i}.tool.json_schema
reasoning_content	metadata.reasoning_content
session_id	session.id
Step messages	llm.input_messages / llm.output_messages
Tool calls	llm.output_messages.{i}.message.tool_calls
Observations	Tool span output.value

Deterministic IDs

Trace and span IDs are derived from the trajectory’s session_id via SHA-256, so uploading the same trajectory twice produces the same trace. This makes uploads idempotent — you can safely re-run without creating duplicates.

Known limitations

Each LLM span includes the full conversation history up to that point as llm.input_messages. For very long sessions (roughly 16+ turns with dense tool calls), this can exceed OpenTelemetry attribute size limits, causing span data to be truncated. This matches the behavior of real-time instrumentors and is a known platform-wide limitation.

API reference

For full parameter documentation, see the API reference.