osmosMind

Data/Agents

Agent Sessions & Tool Trajectories

Session-level data built for agent training. We capture full action sequences, tool calls, and environment-state changes from real dev sessions — including failure, backtracking, and recovery, which is where long-horizon capability comes from.

Coverage

  • Real dev sessions
  • Tool calls
  • Environment state
  • Failure recovery
  • Long-horizon execution

Deliverables

  • Session logs
  • Action sequences
  • State snapshots
  • Result verification
  • Failure causes & trajectory summaries

The pipeline

From source to acceptance

We don't hand-label a pile and ship it. Every category moves through a closed, instrumented loop — generated to a brief, checked by machines, adjudicated by experts, and traceable end to end — but the path each data type takes is its own.

  1. 01

    End-to-End Capture

    Complete sessions are recorded across real and simulated environments, preserving every tool call, observation, and intermediate decision from first prompt to final outcome.

  2. 02

    State Alignment

    Each action is pinned to the environment state it acted on, so the whole trajectory stays faithfully replayable and every step can be re-grounded in its exact context.

  3. 03

    Scripted Verification

    A deterministic check settles the outcome of each consequential action, turning a loose transcript into a sequence of verified results with unambiguous success and failure signals.

  4. 04

    Failure & Recovery Labeling

    Dead ends, backtracks, and recoveries are labeled explicitly, capturing the corrective behaviour that is precisely where durable long-horizon capability is learned.

  5. 05

    Trajectory Summarization

    Each session is distilled into a structured summary of goal, plan, key branches, and recovery paths, making long multi-hundred-turn traces legible and reusable.

  6. 06

    Consistency Review

    A review pass reconciles the summary against the verified steps, ensuring the recorded narrative matches what actually happened in the environment.

  7. 07

    Stratified Sampling

    A final stratified draw across task types, horizons, and failure modes fixes the distribution and guards quality before the trajectories ship.

Every run emits a learning signal that feeds back into the source set — the pipeline tightens itself, batch over batch.

A specimen

See the data itself

One real, trimmed sample from this category — the scenarios it serves, why it matters for training, and the shape of the data as delivered.

Where it’s used

  • Training long-horizon tool-use and planning behaviour
  • Teaching failure recovery and backtracking from real traces
  • Distilling frontier coding-agent behaviour into smaller models

Why it matters for training

Critical

Long-horizon tool use is the current frontier, and real multi-hundred-turn recovery traces are the hardest data to collect at quality.

Notable features

  • 236-turn real sessions
  • Full tool-call + result pairs
  • Failure & recovery captured
  • Replayable state
as-fd9ba3aed9Agent session
Turn 8 of 236 · excerpt
Task

The sandbox directory is empty. Write a local command-line todo tool in Python: a package todo/ (store.py for the data model + in-memory store, cli.py using argparse), supporting the three subcommands add / list / done, with pytest tests and a Makefile. First list a plan with TodoWrite, then implement it, and run make test to confirm everything passes.

  1. Assistant

    List the plan first, then start.

    TodoWrite
    { "todos": [
  { "content": "Create the todo/ package and data model", "status": "in_progress" },
  { "content": "Implement the add/list/done subcommands", "status": "pending" },
  { "content": "Write pytest tests and the Makefile", "status": "pending" }
] }
  2. Result
    Todos have been modified successfully. Continue with the plan.
  3. Assistant

    First check the current state of the working directory.

    Bash
    { "command": "pwd && ls -la" }
  4. Result
    /workspace
total 0
drwxr-xr-x  2 lx staff   64 Jun  3 12:04 .
drwxr-xr-x 18 lx staff  576 Jun  3 12:04 ..
  5. Assistant

    The directory is empty; start writing the in-memory storage layer.

    Write
    { "file_path": "todo/store.py", "content": "from dataclasses import dataclass\n\n@dataclass\nclass Task:\n    id: int\n    text: str\n    done: bool = False\n..." }
  6. Result
  7. Assistant

    Run acceptance once the implementation and tests are written.

    Bash
    { "command": "make test" }
  8. Result
    ===== test session starts =====
collected 6 items
tests/test_todo.py ......                    [100%]
===== 6 passed in 0.04s =====
DONE: 6 tests passing

Need a sample or a custom build?

Tell us your spec and scale — we deliver to order.

Get in Touch
← Back to Data