# Managed agents 101 Canonical: https://mudpie.ai/101/managed-agents/ Updated: 2026-09-20 Curator: Ali Abouelatta (https://mudpie.ai/authors/ali-abouelatta/) The control planes, durable execution, sandboxes, logs and human approvals that let an agent keep working after the first model call ends. **TL;DR:** Managed agents are a systems problem around the model: durable state, isolated compute, tools, secrets, streaming, approvals and recovery. Read Anthropic and OpenAI for the hosted platforms, then Omnara for the open control-plane argument. ## Understand the managed platform These are the current first-party definitions from Anthropic and OpenAI. Compare what each platform owns before deciding what remains application code. ### [Claude Managed Agents overview](https://platform.claude.com/docs/en/managed-agents/overview) Authority: [Anthropic](https://mudpie.ai/authors/anthropic/) **TL;DR:** Claude Managed Agents supplies the harness, environments, sessions, tool execution and durable state around Claude so teams do not rebuild the control plane. **What you'll read:** The core objects, execution lifecycle, security model and the boundary between Anthropic's managed layer and the agent you define. ### [Claude Managed Agents: get to production faster](https://claude.com/blog/claude-managed-agents) Authority: [Anthropic](https://mudpie.ai/authors/anthropic/) **TL;DR:** Anthropic positions managed agents as composable cloud infrastructure for long-running work, not another prompt template or chatbot framework. **What you'll read:** The launch case, agent-environment-session model, examples, observability and the product decisions behind the hosted service. ### [Scaling Managed Agents: decoupling the brain from the hands](https://www.anthropic.com/engineering/managed-agents) Authority: [Anthropic](https://mudpie.ai/authors/anthropic/) **TL;DR:** Separating model reasoning from execution workers lets the platform scale, recover and secure long-running agents without tying identity to one process. **What you'll read:** Anthropic's architecture for orchestration, sandboxes, event streams, state, permissions and production scaling. ### [Introducing the Agents API](https://openai.com/index/introducing-the-agents-api/) Authority: [OpenAI](https://mudpie.ai/authors/openai/) **TL;DR:** OpenAI's Agents API exposes managed cloud agents powered by the Codex harness, with long-running sessions, environments and orchestration behind one API. **What you'll read:** The API object model, environment choices, tool execution, streaming and the relationship between a managed service and the Codex harness. ### [Introducing workspace agents in ChatGPT](https://openai.com/index/introducing-workspace-agents-in-chatgpt/) Authority: [OpenAI](https://mudpie.ai/authors/openai/) **TL;DR:** Workspace agents package repeatable, connected workflows for teams and run them in the cloud through ChatGPT's shared work surface. **What you'll read:** How a business creates, connects, publishes and governs agents that can work across team knowledge and applications. ## Read the open control-plane argument Omnara's founders make the counter-case: durable agent infrastructure should be model-, machine- and harness-independent, with the event log as the stable identity. ### [Omnara introduction](https://docs.omnara.com/introduction) Authority: [Ishaan Sehgal](https://mudpie.ai/authors/ishaan-sehgal/) **TL;DR:** Omnara is an open-source control plane that manages execution and state while letting the builder choose models, tools, machines and deployment. **What you'll read:** The platform objects, event stream, sessions, machines, approvals and the portability promise behind an open managed-agent layer. ### [The Log Is the Agent](https://www.omnara.com/blog/the-log-is-the-agent) Authority: [Ishaan Sehgal](https://mudpie.ai/authors/ishaan-sehgal/) **TL;DR:** An agent's durable identity is the append-only history of inputs, outputs, tool calls and state transitions—not the current model process. **What you'll read:** A systems argument for event sourcing, replay, recovery, portability and observability across models and machines. ### [Serverless Agents](https://www.omnara.com/blog/serverless-agents) Authority: [Ishaan Sehgal](https://mudpie.ai/authors/ishaan-sehgal/) **TL;DR:** If no machine owns the loop, an agent can pause, resume and move across compute while the control plane preserves state and intent. **What you'll read:** The architecture tradeoffs behind decoupled execution, durable queues, resumability and assigning work to ephemeral machines. ### [What Is an Async Agent, Really?](https://www.omnara.com/blog/what-is-an-async-agent-really) Authority: [Kartik Sarangmath](https://mudpie.ai/authors/kartik-sarangmath/) **TL;DR:** Async agents are not just slow requests; they need a durable interaction model for later input, approvals, progress and failure. **What you'll read:** A conceptual teardown of asynchronous work, sessions, user return paths and what a platform must remember after the request ends. ### [The Harness Doesn't Matter](https://www.omnara.com/blog/the-harness-doesnt-matter) Authority: [Kartik Sarangmath](https://mudpie.ai/authors/kartik-sarangmath/) **TL;DR:** Most agent harnesses reduce to a small loop; durable state, tools, permissions and operations create more differentiation than another wrapper around that loop. **What you'll read:** A deliberately contrarian explanation of the harness core and the infrastructure that begins where the simple loop stops. ## Watch the builders explain the hard parts These talks are the shortest path to the operating details: long horizons, memory, recovery, safety and shipping a real managed agent. ### [Ship your first Managed Agent](https://www.youtube.com/watch?v=19HDQ9HppOA) Authority: [Anthropic](https://mudpie.ai/authors/anthropic/) **TL;DR:** A working incident-investigator agent makes the managed platform concrete: define an agent, environment and session, then stream and supervise the run. **What you'll read:** A 37-minute first-party build walkthrough covering setup, tools, execution, events and production-shaped debugging. ### [Claude for Long-Horizon Tasks](https://www.youtube.com/watch?v=9QebvrrY3KY) Authority: [Lance Martin](https://mudpie.ai/authors/lance-martin/) **TL;DR:** Long-horizon reliability comes from the harness and environment around Claude: decomposition, state, feedback, security and recovery. **What you'll read:** An Anthropic engineer's lessons from building agents that work for longer than one turn, presented at AI Engineer. ### [Memory and dreaming for self-learning agents](https://www.youtube.com/watch?v=RtywqDFBYnQ) Authority: [Anthropic](https://mudpie.ai/authors/anthropic/) **TL;DR:** Agent memory needs deliberate consolidation and reflection; saving every token is not the same as learning from experience. **What you'll read:** A first-party session on memory design, offline reflection, durable knowledge and the risks of self-modifying agent behavior. ### [Building more effective AI agents](https://www.youtube.com/watch?v=uhJJgc-0iTQ) Authority: [Anthropic](https://mudpie.ai/authors/anthropic/) **TL;DR:** Reliable agent systems start with simple workflows and add autonomy only when the task and evidence justify the extra uncertainty. **What you'll read:** Anthropic's patterns for prompt chaining, routing, parallelization, orchestrator-workers and evaluator-optimizer loops. ### [Workspace agents](https://openai.com/academy/workspace-agents/) Authority: [OpenAI](https://mudpie.ai/authors/openai/) **TL;DR:** A managed team agent combines instructions, knowledge, tools, publishing controls and repeatable runs inside a shared workspace. **What you'll read:** OpenAI's practical guide to agent anatomy, workflow examples, connections, testing and team deployment in ChatGPT.