Tell Claude What Done Looks Like: Add a Grader Agent to Your Pipeline With Managed Agents Outcomes

Claude Managed Agents Outcomes wires a rubric-graded iterate→grade→revise loop into your pipeline — define success once, and the harness iterates until the output passes.

Most agentic pipelines have a single-shot weakness: the agent runs once, produces output, and you either accept it or retry manually. Anthropic's Managed Agents platform has a better pattern — Outcomes — where you attach a grading rubric to a session and a separate grader agent iterates with the worker until the artifact passes, the iteration cap hits, or you interrupt.

How it works

Send a user.define_outcome event after creating a session:

from anthropic import Anthropic

client = Anthropic()

# Session already created (agent + environment set up separately)
session = client.beta.sessions.create(
    agent=agent.id,
    environment_id=env.id,
    title="Monthly sales report",
)

RUBRIC = """
## Report Rubric
- Contains an executive summary 150 words or fewer
- All figures sourced from the attached CSV (no invented numbers)
- Includes a year-over-year comparison table
- Output is a single .md file
"""

# Wire the grader — agent starts working immediately on receipt
client.beta.sessions.events.send(
    session_id=session.id,
    events=[{
        "type": "user.define_outcome",
        "description": "Generate monthly sales report from attached CSV",
        "rubric": {"type": "text", "content": RUBRIC},
        "max_iterations": 4,  # default 3, max 20
    }],
)

No additional user message is required — the agent starts working as soon as it receives the user.define_outcome event.

Reading evaluation results

Poll or stream span.outcome_evaluation_end events:

session = client.beta.sessions.retrieve(session.id)
for ev in session.outcome_evaluations:
    print(ev.result)       # "satisfied" | "needs_revision" | "max_iterations_reached"
    print(ev.explanation)  # "All 4 criteria met: summary is 142 words, table present..."

The grader runs in a separate context window — it evaluates the output against your rubric cold, without seeing the agent's reasoning. Failures come back as needs_revision with an explanation that feeds directly into the next iteration. The session retains full history, so you can chain a second outcome or continue conversationally after the first one completes.

When to use it

Outcomes work best when "done" is measurable: all tests pass, specific fields are populated, format constraints are met. Anthropic's internal benchmarks showed +8.4% on Word document generation and +10.1% on PowerPoints from simply attaching a rubric. Write specific criteria — "all public functions have type annotations, no bare except clauses" beats "clean code" every time.

Sources: Define outcomes — Claude Platform Docs, Cookbook: verify with outcome grader