Anthropic Details Architecture Behind Claude Agents That Slashed Startup Latency by 90 Percent

Six months after publicly launching Managed Agents on the Claude Developer Platform, Anthropic has published a formal architecture paper documenting the design decisions that make long-running, multi-step AI agents viable in production. The paper, "Scaling Managed Agents with Stable Interfaces," updated October 30, 2026 at anthropic.com/engineering/managed-agents, reveals that decoupling the Claude reasoning layer from its execution environment cut p95 time-to-first-token latency by more than 90 percent — a result that validates the architectural bet at the center of the platform's design.

The Service Behind the Paper

Anthropic first launched Managed Agents in public beta on April 8, 2026, alongside an initial version of the architecture paper written by engineers Lance Martin, Gabe Cemaj, and Michael Cohen. The Managed Agents launch blog describes the service as a hosted agent runtime for tasks that need to run for hours or days without a human in the loop — invoice processing, code review pipelines, multi-step data analysis, and similar autonomous workflows. Early customers named at launch included Notion, Asana, Rakuten, Sentry, and Vibecode.

The platform has expanded steadily since. AWS availability arrived in May 2026, dynamic workflow support entered beta in October, OpenTelemetry-based observability support was added in October, and Google Cloud hosting launched October 22. The October 30 architecture paper update formalizes what six months of production operation has taught the engineering team about making long-horizon agents reliable.

Brain, Hands, and Memory: Three Components That Must Stay Separate

The core architectural insight in the paper is that a production agent system comprises three fundamentally distinct components that should not share state: the Session, the Harness, and the Sandbox.

The Session is an append-only, durable event log stored outside any individual container. It records everything the agent has done across its entire lifetime. Developers interact with it through a getEvents() call that returns positional slices, not a streaming context buffer. Critically, the session is not Claude's context window — it is an external store that survives container restarts, model upgrades, and network interruptions.

The Harness — what the paper calls "the brain" — is the stateless component that runs Claude's decision loop. It queries the session, selects which events to pass into Claude's context window (not all of them, and not verbatim), invokes Claude for a decision, and then delegates execution to the sandbox. The harness never holds persistent state between invocations, which makes it trivially scalable: any available harness instance can resume any session.

The Sandbox is the execution environment — "the hands." It runs tools, executes code, manages file systems, and calls external APIs. The paper describes it as "cattle, not pets": when a container becomes unhealthy or finishes its work, it is terminated rather than repaired. A new sandbox is provisioned on demand via provision({resources}), receives a fresh state, and picks up where the session log indicates the last successful step completed.

Why Session State Must Live Outside Claude's Context Window

The paper introduces the term "context anxiety" to describe a behavioral pattern discovered in Claude Sonnet 4.5: as the model's context window filled up during a long task, it would begin prematurely summarizing and wrapping up, treating the approaching context limit as a signal that the task should conclude. Engineers added context reset mechanisms to the harness to address this — but those resets introduced their own problems when Anthropic updated the underlying model. Claude Opus 4.5 did not exhibit the same context anxiety, so the resets became unnecessary overhead that degraded performance without benefit.

This experience directly motivated the stable-interface design principle at the paper's core: the contract between the harness, the session, and the sandbox must outlast any specific model generation. The paper draws an analogy to POSIX's read() system call, which has survived unchanged from 1970s disk hardware to modern solid-state storage because it abstracts away implementation details behind a stable interface. When Anthropic ships a new Claude version, existing harness code should not require modification. The session-harness-sandbox interface contract — execute(name, input) → string at the harness-sandbox boundary — is designed to remain stable as models are upgraded beneath it.

Credential Security Without Harness Visibility

A notable security design documented in the paper is how OAuth credentials are handled. Rather than storing authentication tokens in the harness or passing them through the execution environment, Anthropic uses a vault-and-proxy model: tokens are stored in an external credential vault and accessed by the sandbox through an MCP (Model Context Protocol) proxy. The harness never receives or holds OAuth credentials directly. For git-based workflows, authentication tokens are baked into the repository clone at initialization time rather than maintained in runtime state, avoiding credential exposure through session event logs or harness memory.

Many Brains, Many Hands: Agent Scaling

Because both the harness and the sandbox are stateless (the session holds all durable state), scaling to multiple parallel agents requires no coordination layer between harness instances. Each harness awakens a session via wake(sessionId), processes the current event queue, dispatches to sandboxes, and emits results back to the session log via emitEvent(id, event). Multiple harness instances can operate against different sessions simultaneously without shared infrastructure. Sandboxes are provisioned with explicit resource requirements — provision({resources}) — making it straightforward to allocate heavier compute for tasks that require it and minimal resources for tasks that do not.

The API surface is intentionally minimal: wake, getSession, emitEvent, provision, and execute. All requests require the beta header managed-agents-2026-04-01. Pricing follows a session-time model at $0.08 per session-hour, plus standard Claude token costs.

Where Anthropic Sits in the Emerging Agent Hosting Market

Anthropic is not operating alone in this space. OpenAI launched its Agents API in public beta on September 10, 2026, built on the Codex harness. OpenAI's pricing model differs: the Agents API charges only for model token consumption and tool usage, with no separate session-hour fee. Both platforms share the architectural pattern of externalized session state and isolated sandboxes — the managed-hosting-for-agents approach appears to be converging as an industry standard for enterprise AI deployment.

Google offers agent infrastructure through Vertex AI Agent Builder, and Anthropic's Managed Agents are now available on Google Cloud following the October 22 launch, creating a cross-cloud option for organizations already using GCP infrastructure.

The competitive distinction Anthropic is making with this paper is less about pricing than about architectural transparency. By formalizing the stable-interface commitment in a public document, Anthropic is signaling to platform engineers that the session-harness-sandbox contracts are not subject to arbitrary change and that agent code written against today's API will survive the next model release.

What Comes Next for Production Agent Infrastructure

The Managed Agents service remains in public beta. General availability has not been announced. The feature additions through October — dynamic workflows, observability via OpenTelemetry, multi-cloud reach — suggest the platform is maturing toward GA, but Anthropic has not confirmed a timeline.

The more significant near-term question is whether the stable-interface abstraction will hold as AI models change more rapidly. The paper's argument is architectural and somewhat philosophical: just as operating systems decoupled application code from hardware details, agent platforms should decouple task logic from model-specific behavior. Whether that commitment remains stable when the next generation of Claude introduces capabilities that don't map cleanly onto the current interface contract is the test this design will face.

For now, the architecture paper offers the clearest public documentation yet of how Anthropic thinks production agents should be built — and the performance data suggests that thinking, at least across six months of real workloads, has produced measurable results. The full paper is available at anthropic.com/engineering/managed-agents.