import agentpile as ap

run = next(ap.load("who_and_when"))   # one of 13 datasets by name, or a trace file by path
ap.show(run)                          # every step, call and result, with the source's labels beside them

Public datasets store agent runs in their own schemas (e.g., chat messages in tau-bench, OpenTelemetry spans in TRAIL, raw log text in MAST), and each analysis tool expects its own input. agentpile sits between the two. It reads datasets, trace files and live runs into ATIF v1.7 records that keep what the source recorded. You can then use those records with pandas, networkx or a tracing UI such as Phoenix. It downloads each dataset from its publisher at a pinned revision and redistributes nothing.

1. What You Can Do

1.1 Open Any Run in a Tracing UI

tau-bench stores plain chat logs. ap.send_otlp posts a run as OTLP spans to Arize Phoenix, Jaeger, MLflow or any OTLP/HTTP collector.

run = next(ap.load("tau_bench", limit=1))
ap.send_otlp(run, "http://localhost:6006/v1/traces")   # Phoenix; Jaeger listens on :4318

A run whose source has no timestamps gets synthetic ones that only order its spans. The durations a tracing UI shows for such a run are therefore not runtime measurements.

1.2 Compare Failures Across Datasets

Who&When blames a step, TRAIL marks error spans, and MAST flags 14 failure modes per run. One table holds all three (script).

# TRAIL is gated: accept its terms on Hugging Face and log in first (section 3)
runs = [run for name in ("who_and_when", "trail", "mast") for run in ap.load(name)]
labels = ap.to_pandas(runs, "labels")   # run_id, name, value, target_kind, step, span, ...

1.3 Follow Values Through a Run

Which earlier result does each call reuse? When a call reuses an id-like value from an earlier result, the value_flow rule makes it depend on the call that produced that result. Such values include reservation ids, product ids and dates. The links are inferred from the recorded values (script).

deps = ap.derive_deps(run, rules=["value_flow"])        # each entry names its rule and version
graph = ap.to_networkx(run, rules=["value_flow"])       # events as nodes, dependencies as edges

2. Quick Start

pip install agentpile                    # the core: standard library only
pip install "agentpile[hf,view]"         # adds dataset downloads and the colored ap.show

Extra	Adds
hf	downloads of the Hugging Face datasets
view	colored ap.show output (rich); without it the view is plain text
pandas, graph	ap.to_pandas and ap.to_networkx; ap.table gives the same rows with the core alone
otel	protobuf for ap.send_otlp to a receiver that takes only protobuf, such as Phoenix (Python 3.10+)
pydantic-ai, openai-agents	ap.capture for Pydantic AI (Python 3.11+) and the OpenAI Agents SDK (Python 3.10+)

A conversation in the OpenAI chat format needs only the core:

import agentpile as ap

messages = [
    {"role": "user", "content": "Is it raining in Paris?"},
    {"role": "assistant", "content": None, "tool_calls": [{"id": "call_1", "type": "function",
        "function": {"name": "get_weather", "arguments": "{\"city\": \"Paris\"}"}}]},
    {"role": "tool", "tool_call_id": "call_1", "content": "light rain, 14 C"},
    {"role": "assistant", "content": "Yes, light rain."},
]
run = next(ap.load(messages))             # the format is detected, and an ambiguous input is an error
for event in run.events():                # messages, tool calls and tool results in source order
    print(event.kind, event.role, event.content)

for dep in ap.derive_deps(run, rules=["tool_call_id"]):   # inferred only when you ask, by a named rule
    print(dep["rule"], "step", dep["src"]["step"], dep["src"]["call"], "-> step", dep["dst"]["step"])

run.to_json("run.json")                   # ATIF v1.7 JSON; run.to_otlp("run.otlp.json") writes OTLP spans

message user Is it raining in Paris?
message agent None
call agent {'city': 'Paris'}
result tool light rain, 14 C
message agent Yes, light rain.
tool_call_id step 2 call_1 -> step 3

The second line is a stand-in made visible. ATIF does not permit the null in the source's first assistant message. agentpile therefore stores an empty string in the record and notes that the source held null, and the event reports None. Rule tool_call_id gives the dependency: it makes the next agent step after a result depend on the call that produced it, here the final answer (step 3) on call_1 in step 2.

A dataset loads by name (ap.datasets.names() lists them), and a trace file by its path (ap.readers.names() lists the formats):

for run in ap.load("who_and_when", limit=5):   # fetches only the files these five runs come from
    print(run.id, run.labels)                  # the labels the reader emits, each with its target
run = next(ap.load("my_trace.json"))           # ATIF, OpenAI messages, OTLP JSON or a tool list

3. Datasets

Each name loads one dataset at the revision shown, and a full load is checked against the run count in its manifest. Harbor's ATIF v1.7 models (Harbor 0.22.0) accept every trajectory checked: the first 3,000 runs of each dataset, or all of a smaller one. tools/dataset_table.py regenerates this table.

Name	Source at pinned revision	Runs	Source annotation levels	License on the card
who_and_when	Kevin355/Who_and_When @ 59b9fcb	184	run, step	not declared
mast	mcemri/MAST-Data @ 95118ac	1,642	run	cc-by-4.0
trail	PatronusAI/TRAIL @ b424ce6	148	run, span	mit
tau_bench	AgentSuite/tau-bench-trajectories @ 382e57d	4,950	run	not declared
tau2_bench	AgentSuite/tau2-bench-trajectories @ 07264f1	8,340	run	not declared
agent_reward_bench	McGill-NLP/agent-reward-bench @ b6d17e6	1,302	run	not declared
agentdojo	sungjuncho/agentdojo-trajectories @ 43df1b8	15,095	call, run, span	not declared
swe_rebench_openhands	nebius/SWE-rebench-openhands-trajectories @ 3545538	67,074	run	cc-by-4.0
swe_agent_nebius	nebius/SWE-agent-trajectories @ 68195a1	80,036	run, step	cc-by-4.0
swe_gym_openhands	SWE-Gym/OpenHands-Sampled-Trajectories @ baf3a4e	6,055	run	not declared
swe_smith	SWE-bench/SWE-smith-trajectories @ 08e109b	24,100	run, step	mit
adp	neulab/agent-data-collection @ 31a76bf	200,914	run	not declared
catchbench_pre	yzhao062/catchbench @ 30a516c	1,187	run, tool	MIT; rows name their upstream license

Notes on the table and on single datasets

• Annotation levels. The column lists the levels at which each source annotates its runs, as its manifest records them. Readers turn part of these annotations into run.labels: run labels for every dataset except ADP, plus step and agent labels for Who&When, span labels for TRAIL and tool labels for CatchBench PRE. The rest stay elsewhere in the record (e.g., ADP's results in run.outcome, SWE-smith's patch in run.dataset).
• TRAIL is gated on Hugging Face: accept its terms on the dataset page with the account whose token you use.
• AgentRewardBench run files total about 30 GB, because each step keeps the browser state the agent saw. limit= fetches only the run files it reads.
• SWE-smith loads its tool split. 14,984 of its 24,100 rows write actions as text and load as plain messages, and message_format="native" returns only the 9,116 rows with structured calls.
• ADP loads the 14 configurations whose records pass validation and a count check, and the manifest lists the rest. The loader follows ADP's own ADP-to-ATIF converter with two changes, both to keep the source's roles. An environment observation that answers no call becomes a user step whose role is environment, where ADP's converter makes it an empty agent step. A call's result keeps the observation's own source, where ADP's converter relabels user as environment.

4. Add a Dataset

Have agent trajectories or a trace format that agentpile does not load yet? Open an issue with a link and we write the loader, the manifest and the tests. A pull request with a loader is welcome too (see Extending).

5. Capturing a Running Agent

ap.capture records a run of a Pydantic AI agent or an OpenAI Agents SDK agent into the same record, with the agent's configuration attached:

pip install "agentpile[pydantic-ai]"     # or "agentpile[openai-agents]"

with ap.capture(agent=my_agent) as cap:
    my_agent.run_sync("Book a flight from LAX to SFO.")
cap.run            # the captured run
cap.otlp           # the same spans as an OTLP JSON document
ap.config_from(my_agent)   # the configuration alone: instructions, model, tools with schemas, handoffs

Capture is tested with Pydantic AI 2.58.0 and 2.59.0 and with the OpenAI Agents SDK 0.24.0 through openinference-instrumentation-openai-agents 2.5.3. agentpile checks the version of the instrumentation that wrote the spans, which is Pydantic AI itself or, for the Agents SDK, OpenInference. With any other instrumentation version cap.run raises an error, and ap.capture(my_agent, allow_untested_version=True) builds the run anyway and marks it source.dialect_tested false.

Which runs a capture records

A capture uses its own tracer provider and never sets the global one. Observers the host program already had keep receiving every span, and the agent's original instrumentation setting is restored when the block exits, including on an exception. Everything a capture records happens while the block is open and only for the agents passed to capture(). Runs of other agents in the same window are not recorded. What counts as a selected agent's run depends on the framework. For the Agents SDK, a trace in which a selected agent takes part (it starts the run, receives a handoff, or is called as a tool) is recorded whole, other participants included. Because the SDK names agents only by name in its spans, an unselected agent that shares a selected agent's name is also recorded. For Pydantic AI, each selected agent's own runs are recorded: an agent that a selected agent calls from a tool is recorded only when it is selected too. Pass every agent whose delegated work you want.

6. Reference

6.1 The Record

A trajectory record is an ATIF v1.7 trajectory, the Agent Trajectory Interchange Format of the Harbor evaluation framework. What ATIF has no field for sits under extra["agentpile"]: where the run came from, the source's labels with their targets, the agents and tools with their permissions, and the source value behind each message or call stand-in. A configuration with no execution (e.g., an MCP tool list) is a config record (agentpile-config/0) with the same agents, tools and labels at its top level and no steps.

Loading never adds what the source did not record: it keeps any dependencies the input states and infers none. The few pairings a reader has to infer carry paired_by with the name of the rule, and ap.derive_deps(run, rules=[...]) infers dependencies between steps when you ask.

What the extension holds, and how inference and export work

• Source and identity: dataset, revision, file, record key and reader version, enough to trace every run back to the row it came from.
• Labels with targets: each label points at the run, an agent, a step, a call, a span or a tool, with the label's origin (expert, LLM judge, benchmark).
• Agents and tools: each agent's instructions, model and tools, and each tool's original definition, its permission, and where that permission came from.
• Stand-ins made visible: where ATIF requires a value the source did not have (a non-null message, an object of arguments, a call id, an image part it cannot hold), the record keeps the source value and the event view reports it. A missing value, a null and an empty string therefore stay distinct. A call whose source arguments were neither a JSON object nor null keeps them in arguments_raw and carries arguments_state: "unparsed", and in run.to_otlp() its model span and its tool span both show that source value. For any other call the model span shows the parsed arguments, and the tool span shows the source text where the reader kept it.
• Source order: run.events() follows the order the source recorded, which differs from ATIF where ATIF stores a result on the step that issued the call (a delegated agent's steps can come in between). A reader whose references are not integers (e.g., TRAIL's span addresses) stores an integer sort position beside each reference (source_pos, result_pos), and the reference itself keeps pointing back into the source.

CatchBench PRE rows load as config records, and so do tool definitions: an MCP tools/list result, an OpenAI tool list and an Anthropic tool list (readers mcp-tools-list, openai-tools and anthropic-tools), each definition kept verbatim. An MCP tool's permission comes from its annotations when it states them and is unknown otherwise.

Inferred pairings include a result matched to its call by position, a TRAIL history call placed on the latest agent step, and a TRAIL TOOL span tied to its call by a unique match. Every dependency from ap.derive_deps names the rule and version that produced it, and run.with_deps(rules=[...]) returns a copy whose steps carry those entries, which to_otlp() then writes as span links. The run you loaded is not changed.

run.to_otlp() writes a run as OTLP JSON spans in the layout of Harbor's ATIF-to-OpenTelemetry converter. A run loaded from a trace writes each of its source spans once, and a telemetry tool therefore counts the same model and tool operations as in the original trace. The spans agentpile adds around them (the trajectory, user and system steps) are CHAIN spans, not model or tool operations. Reading an OTLP trace whose instrumentation version differs from the tested ones raises an error unless you pass allow_untested_version=True, which marks the record as read from an untested version.

6.2 Extending

A reader is any object with sniff(source) -> bool and read(source) -> Iterator[Run]. Register one for a session with ap.register("my-format", reader), or ship it in a package through the entry-point group agentpile.readers (datasets and writers have agentpile.datasets and agentpile.writers). Pass reader="..." to ap.load to skip detection.

6.3 Downloads and Cache

Downloaded files live in one place, ~/.cache/agentpile/<dataset>/<revision>/ (set AGENTPILE_CACHE to move it), and never in the shared Hugging Face cache. A load with limit= fetches only the files its runs come from, and a later full load fetches the rest. Until a full download completes, each limited load still asks the hub for its file listing. ap.cache_info() lists what is there and ap.clear_cache() removes it.

7. Status

Alpha. Version 0.1.0 covers the 13 datasets above, each checked against a manifest of pinned counts, the ATIF, OpenAI-style message, OTLP JSON, tool-definition and agentpile config formats, and capture for Pydantic AI and the OpenAI Agents SDK. The record format may still change before 1.0, and the changelog lists every change.

8. License

Apache-2.0 for the code. Each dataset keeps its publisher's terms. Several declare no license on their dataset card, and agentpile records what each card says in its manifest. The tests use synthetic fixtures in each dataset's file layout and copy no dataset content. The name follows the habit of calling a mixed collection a pile and has no connection to EleutherAI's The Pile.

back to top