Define outcomes

An outcome tells the session what the end result should look like and how to measure its quality. The agent works toward that target, self-evaluating and iterating until the outcome is met.

When you define an outcome, the harness automatically provisions a grader to evaluate the artifact against a rubric. The grader uses a separate context window to avoid being influenced by the main agent's implementation choices.

The grader returns an explanation summarizing which criteria passed or failed, or confirming that the artifact satisfies the rubric. That feedback is handed back to the agent for the next iteration.

Create a rubric

A rubric is a markdown document describing per-criterion scoring. The rubric is required.

Example rubric:

# DCF Model Rubric

## Revenue Projections
- Uses historical revenue data from the last 5 fiscal years
- Projects revenue for at least 5 years forward
- Growth rate assumptions are explicitly stated and reasonable

## Cost Structure
- COGS and operating expenses are modeled separately
- Margins are consistent with historical trends or deviations are justified

## Discount Rate
- WACC is calculated with stated assumptions for cost of equity and cost of debt
- Beta, risk-free rate, and equity risk premium are sourced or justified

## Terminal Value
- Uses either perpetuity growth or exit multiple method (stated which)
- Terminal growth rate does not exceed long-term GDP growth

## Output Quality
- All figures are in a single .xlsx file with clearly labeled sheets
- Key assumptions are on a separate "Assumptions" sheet
- Sensitivity analysis on WACC and terminal growth rate is included

Pass the rubric as inline text on user.define_outcome (see Create a session with an outcome), or upload it through the Files API for reuse across sessions.

import time
from pathlib import Path

from anthropic import Anthropic

client = Anthropic()

RUBRIC = """# DCF Model Rubric

## Revenue Projections
- Uses historical revenue data from the last 5 fiscal years
- Projects revenue for at least 5 years forward

## Output Quality
- All figures are in a single .xlsx file with clearly labeled sheets
"""
Path("/tmp/rubric.md").write_text(RUBRIC)

rubric = client.files.upload(file=Path("/tmp/rubric.md"))
print(f"Uploaded rubric: {rubric.id}")

Create a session with an outcome

The following examples create a session for an existing agent and environment (both created separately), then send a user.define_outcome event. The agent begins work immediately. No additional user message event is required.

# Create a session
session = client.beta.sessions.create(
    agent=agent.id,
    environment_id=environment.id,
    title="Financial analysis on Costco",
)

# Define the outcome — agent starts working on receipt
client.beta.sessions.events.send(
    session_id=session.id,
    events=[
        {
            "type": "user.define_outcome",
            "description": "Build a DCF model for Costco in .xlsx",
            "rubric": {"type": "text", "content": RUBRIC},
            # or: "rubric": {"type": "file", "file_id": rubric.id},
            "max_iterations": 5,  # optional; default 3, max 20
        }
    ],
)

Outcome events

Progress on an outcome-oriented session is surfaced on the events stream.

• agent.* events (such as messages and tool use) show progress toward the outcome.
• span.outcome_evaluation_* events are only emitted for outcome-oriented sessions and show the number of iteration loops and the grader's feedback process.
• You can also send user.message events to an outcome-oriented session to direct the agent's work as it progresses, but it isn't required: the agent works toward the outcome on its own, iterating until it succeeds or runs out of iterations.
• A user.interrupt event pauses work on the current outcome and marks the span.outcome_evaluation_end.result as interrupted, allowing you to kick off a new outcome.
• After the final outcome evaluation, the session can be continued as a conversational session, or a new outcome can be started. The session retains history of the prior outcome.

Define outcome user event

This is the event you send to initiate an outcome. It is echoed back on receipt, including a processed_at timestamp and outcome_id.

{
  "type": "user.define_outcome",
  "description": "Build a DCF model for Costco in .xlsx",
  "rubric": { "type": "file", "file_id": "file_01..." },
  "max_iterations": 5
}

Outcome evaluation start

Emitted once the grader starts an evaluation over one iteration loop. The iteration field is a 0-indexed revision counter: 0 is the first evaluation, 1 is the re-evaluation after the first revision, and so on.

{
  "type": "span.outcome_evaluation_start",
  "id": "sevt_01def...",
  "outcome_id": "outc_01a...",
  "iteration": 0,
  "processed_at": "2026-03-25T14:01:45Z"
}

Outcome evaluation ongoing

Heartbeat emitted while the grader runs. The grader's internal reasoning is opaque: you see that it's working, not what it's thinking.

{
  "type": "span.outcome_evaluation_ongoing",
  "id": "sevt_01ghi...",
  "outcome_id": "outc_01a...",
  "iteration": 0,
  "processed_at": "2026-03-25T14:02:10Z"
}

Outcome evaluation end

Emitted when an outcome evaluation cycle ends: after the grader finishes evaluating one iteration, or when the session is interrupted while an outcome is active. The result field indicates what happens next.

{
  "type": "span.outcome_evaluation_end",
  "id": "sevt_01jkl...",
  "outcome_evaluation_start_id": "sevt_01def...",
  "outcome_id": "outc_01a...",
  "result": "satisfied",
  "explanation": "All 12 criteria met: revenue projections use 5 years of historical data, WACC assumptions are stated, sensitivity table is included...",
  "iteration": 0,
  "usage": {
    "input_tokens": 2400,
    "output_tokens": 350,
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 1800
  },
  "processed_at": "2026-03-25T14:03:00Z"
}

Check outcome status

You can either listen on the event stream for span.outcome_evaluation_end, or poll GET /v1/sessions/{session_id} and read outcome_evaluations[].result. Until an evaluation completes, result reports pending, running, or evaluating:

session = client.beta.sessions.retrieve(session.id)

for outcome in session.outcome_evaluations:
    print(f"{outcome.outcome_id}: {outcome.result}")
    # outc_01a...: satisfied

Retrieve deliverables

The agent writes output files to /mnt/session/outputs/ inside the sandbox. To retrieve them, list files through the Files API with the session ID as the scope_id, then download them by ID. Filtering by scope_id requires the managed-agents-2026-04-01 beta header on the list request, so the SDK and CLI examples make that call through the beta namespace and pass the header explicitly. Files appear in the list shortly after the agent finishes writing them, sometimes a few seconds after the session goes idle. If a file you expect is not listed yet, list again after a short delay; once it appears in the list, its upload has finished.

# List files produced by this session
# scope_id filtering requires the managed-agents beta on the files request
files = client.beta.files.list(scope_id=session.id, betas=["managed-agents-2026-04-01"])
for file in files:
    print(file.id, file.filename)

# Download a file
if files.data:
    content = client.files.download(files.data[0].id)
    content.write_to_file("/tmp/output.txt")

Next steps