Managed Agents: who runs the loop, a 45-minute workshop

With JavaScript off, the slides and their notes read as one page; the figures and the presses need JavaScript.

Managed Agents: who runs the loop

github.com/brokenairplane/managed-agent-session

Session goal

  • Engineers, and anyone new to agents
  • Evaluate what parts of an agent run on the developer's side or Anthropic's side

Today we work in the lab. Coding is not required: everyone is welcome to stay.

Setup · runs while the talk goes on

Run these three lines now

  1. git clone https://github.com/brokenairplane/managed-agent-session
  2. cd managed-agent-session
  3. python3 lab/setup_lab.py --no-key

While setup runs, open lab/index.html in a browser.

In the lab: the loop the sandbox the wire call or send return or stream

The page at 2am

A page lights a phone at night, sent to an agent whose first reply shows beside the Claude spark, and three thoughts: data, approval, costthe pagecheckout p99 ×10four deploys since 22:00sendthe agent's first replyReading the four deploys…Diff in checkout-api 2.14.0touches the cart cacheRunning: get_recent_deploysdataData that may notleave our networkapprovalWho approves a commandbefore it runs?costWhat does an hourof this cost?no one at the keyboard for the next hour

The request: an agent that handles the first hour of working out what broke, with nobody at the keyboard.

Name one job from your own team's work that an agent could run unattended.

Managed Agents in context

Click a word to see where the thing the word names sits on the map.

What do you want to own?

* the wire: the network boundary. Between the app and Anthropic only API calls and streamed events cross it.

Demo: the 2am triage agent's first send

Managed Agents, cloud sandbox

solid: a send or a request dashed: an event or a result back dotted: inside Anthropic's side

Skipped the setup?

  1. git clone https://github.com/brokenairplane/managed-agent-session
  2. cd managed-agent-session
  3. python3 lab/setup_lab.py --no-key

Already ran them? Open lab/index.html in a browser.

In pairs · 4 minutes

Task: Who answers a pause

  1. Open lab/index.html in a browser; it opens on Task: Who answers a pause
  2. Click Show the diagram, drag Replay step to 1, then click  Next step for each step, naming it a send, a watch or an answer
  3. In bash runs only after, select a person (always_ask) so bash waits for approval; predict which step moves, then click  Next step again

In the lab: the loop the sandbox the wire call or send return or stream

Lab: Who answers a pause, live

  1. The lab, open on Task: Who answers a pause
  2. Click Next step to walk the run, one step at a time

Pick what the on-call team stops owning

Loop run bySandbox run by
developerdeveloper
developerdeveloper
Anthropicdeveloper
AnthropicAnthropic

Which way runs the loop in your own process, using Anthropic's library? Choose one.

Lab · 25 minutes

Task: Where the loop lives

  1. Open lab/index.html in a browser
  2. Open lab/index.html in a browser
  3. Select Task: Where the loop lives
  4. Select a job
  5. Drag the loop and sandbox boxes to the side that runs each
  6. Click Check to see the result

This task comes after the pairs' Task: Who answers a pause.

Lab: Where the loop lives, live

  1. The lab, open on Task: Where the loop lives
  2. Drag the boxes the way a group placed them, then click Check

With Managed Agents, the developer's app sends events, watches them, and answers the agent; Anthropic runs the loop and the sandbox.

Questions?github.com/brokenairplane/managed-agent-session[handle]

Lab progress stays in the browser until the learner copies a summary

Managed Agents lab

Feedback preview

{
  "lab_version": 1,
  "day": "2026-12-03",
  "task_reached": 3,
  "tasks_made": ["d1", "pauses"],
  "misses_per_row": { "pause:bash": 2 },
  "second_miss_hint_shown": true,
  "time_on_task": {
    "task_1": "1 to 5 min",
    "task_2": "5 to 15 min",
    "task_3": "under 1 min",
    "task_4": "under 1 min"
  }
}
  1. Click  Feedback: preview what would be shared to review the summary
  2. Click Copy to take the summary out of the browser and share it

Which is the highest that is true for you now? Choose one.

After the talk

Frequently asked questions

One slide per question, listed on the next slide.

Which step of the triage run does each setting change? Click a run

Loading the four recorded runs

Which settings decide what the agent may do?

  1. Create the agent: model, prompt, tools
  2. Create the environment: where tools run
  3. Start the session: one run
  4. Send a message

Click a setting to see where the setting sits on the map.

What do I actually create, in what order?

Agent, then environment, then a session

What does a session cost?

Tokens, plus $0.08 per running session-hour

A budget sets a ceiling on the session's cost at published prices; at the ceiling the session goes idle and keeps its history

Is there a beta header, and do I set the header?

Required on every call; the SDKs set it

Are custom tools handled through MCP?

No: the app runs a custom tool

What permission policies exist, and what are the defaults?

always_ask: bash waits for the app's yes

Custom tools: the app decides, no policy

Is auto a human-in-the-loop checkpoint?

No: the server judges each call

My session went idle. How do I know why?

Read the idle event's stop reason

How do I show text while the agent writes?

Event deltas give a live preview

What network access does an environment get by default?

Unrestricted, unless the request says otherwise

How do I give the agent files?

Upload, then mount: read-only copies

How do I stop an unattended run from overspending?

Set a session budget at creation

Agent SDK, Managed Agents, or the Messages API: how do I choose?

The SDK's loop runs in the developer's process

Loop run bySandbox run by
developerdeveloper
AnthropicAnthropic

Self-hosted environment: where does the sandbox run?

A self-hosted sandbox moves the sandbox, not the loop

Loop run bySandbox run by
Anthropicdeveloper
AnthropicAnthropic

Does limited networking also govern web search and web fetch?

Yes, under limited networking

If I update the agent, do running sessions change?

Not if the session pins a version

What happens on long runs that exceed the context window?

The service compacts the history

Can agents delegate to other agents or run on a schedule?

Yes: threads, workflow runs, schedules

Can we use Managed Agents under Zero Data Retention or HIPAA?

Hosted: not eligible. Self-hosted Managed Agents: ZDR exception

Transcript kept by
developer
Anthropic
Anthropic

What are the rate limits at scale?

300 a minute to create, 1,200 to read

Is Managed Agents on AWS or Google Cloud, and how does Bedrock AgentCore compare?

On both, with some differences

How do I know the agent is good enough to run unattended?

Evals before launch, many trials

What should production monitoring look like?

Alert on idle, budget and termination

Can learners use their Claude subscription instead of API billing?

Max and Team plans include API credits