Anthropic Ships Five Updates Making Claude Managed Agents Easier to Run in Production

Anthropic adds streaming deltas, per-session agent overrides, new webhook events, and a Console observability tab to Claude Managed Agents

·
·
·
Read6 min
TypeNews
TopicAgents · Api
  • 5 new features shipped for Claude Managed Agents: streaming event deltas, per-session agent overrides, new webhook event types, reverse pagination, and credential injection scoping.
  • Agent Overrides let you start a session from a stored agent and swap the model, system prompt, tools, MCP servers, or skills for just that session -- no new agent version needed.
  • Streaming deltas now emit incremental content chunks during a session turn, enabling real-time UI updates without waiting for a full turn to complete.
  • New Observability tab in the Claude Console surfaces session-level token usage and tool call metrics without custom instrumentation.
  • Reverse pagination lets you jump directly to the tail of a long session event log, useful when debugging sessions with hundreds of events.
  • All features are available now in public beta; pricing remains standard token rates plus $0.08 per session-hour.

Claude Managed Agents just got a meaningful quality-of-life update. Anthropic shipped five new capabilities that close some of the most common friction points when running agents in production: streaming session event deltas, per-session agent overrides, new webhook event types, reverse pagination, and credential injection scoping. None of these are headline model improvements, but together they make the platform substantially easier to wire into real applications.

The platform these updates sit on

Claude Managed Agents provides the harness and infrastructure for running Claude as an autonomous agent. Instead of building your own agent loop, tool execution, and runtime, you get a fully managed environment where Claude can read files, run commands, browse the web, and execute code securely. Secure sandboxing, session persistence, credential isolation, error recovery, and observability are all handled -- engineering teams previously dedicated 4-8 senior engineers for 3-6 months before a single agent reached production. Managed Agents eliminates that entire layer.

You create a persisted agent config covering model, system prompt, tools, MCP servers, and skills, then start sessions that reference it. The session streams events back to you; you send user messages and tool results in. An agent is a persisted, versioned config -- every update creates a new immutable version, and sessions pin to a version at creation time. This lets you iterate on the agent without breaking sessions already running, roll back if a change regresses, and A/B test versions side-by-side.

What just shipped

The five additions announced in this update each target a specific gap in the developer workflow:

  • Streaming session event deltas: Previously, you had to poll or reconnect to get updated event state. Now the event stream sends incremental deltas as they happen, which is essential for building responsive UIs on top of long-running agent sessions. These are raw Anthropic API streaming events -- text deltas, tool input chunks -- that let you display output in real time as Claude generates it.
  • Per-session agent overrides: This is arguably the most useful addition for teams running agents at scale. You can now start a session from a stored agent definition and override specific parts of its config for just that one session -- swap the model, system prompt, tools, MCP servers, or skills -- without creating a new agent version. This is perfect for A/B testing prompts or routing different user tiers to different models without proliferating agent configs.
  • New webhook event types: Anthropic previously announced webhooks for Managed Agents. You can define an outcome, let the agent run, and get notified by a webhook when it's done. The addition of webhooks turns Managed Agents from something you watch into something you wire up. The new event types extend what you can subscribe to, enabling more granular triggers in your downstream systems.
  • Reverse pagination: When listing session events, you can now paginate backwards through history. This matters when you have long-running sessions with thousands of events and you only care about the most recent ones -- previously you had to page through from the beginning.
  • Credential injection scoping: Vaults hold environment variable credentials for non-MCP services (CLIs, SDKs, direct API calls) -- substituted at egress, never visible in the sandbox. The new scoping controls let you specify exactly which credentials are injected for a given session, rather than attaching an entire vault.

The observability tab

Alongside the API updates, Anthropic added a new Managed Agents Observability tab in the Claude Console. This gives you session-level metrics including input and output token use and tool usage, without having to instrument your own logging pipeline. For teams that are still in the early stages of understanding what their agents are actually doing, this is a low-friction starting point.

The official docs describe the broader Console view: the production pattern registers a webhook in the Console that fires on session.status_idled, which is the signal that the agent is either done or waiting on a tool result. The new observability tab sits alongside this, surfacing the token and tool data you need to understand cost and behavior at a glance.

How to use it

You need an Anthropic API key with Managed Agents access. There's no waitlist right now -- any existing key works in public beta. Pricing is standard Claude API token rates plus $0.08 per session-hour. The SDK sets the required managed-agents-2026-04-01 beta header automatically.

To try the new agent overrides, you pass an overrides object when creating a session. Here's a minimal example showing how you'd swap the model for a single session:

makefile
import anthropic
client = anthropic.Anthropic()
session = client.beta.sessions.create(
    agent=AGENT_ID,
    environment_id=ENVIRONMENT_ID,
    # Override just the model for this session
    overrides={
        "model": {"id": "claude-sonnet-4-5"},
        # system, tools, mcp_servers, or skills can also be overridden
    }
)

For the new streaming deltas, the pattern is the same event stream you already use -- the difference is that assistant message events now emit incremental content chunks rather than waiting for a full turn to complete before sending anything.

Who this is for

These updates are most valuable if you're already running Managed Agents in production and hitting specific walls:

  • Teams building chat-like UIs on top of agents will benefit most from streaming deltas -- waiting for a full turn to complete before rendering anything is a poor user experience.
  • Teams with a single stored agent that needs to behave differently for different users (different tiers, different contexts) will find per-session overrides significantly cleaner than maintaining multiple agent configs.
  • Teams with complex event-driven pipelines will benefit from the expanded webhook event types, which reduce the need to poll for state.
  • Anyone debugging long sessions will appreciate reverse pagination -- jumping to the tail of a 500-event session log is now a single API call.

Managed Agents is Claude-specific hosted infrastructure -- you're trading flexibility for operational simplicity. Because of this, Managed Agents is not currently eligible for Zero Data Retention or HIPAA Business Associate Agreement coverage. Those constraints haven't changed with this update, but for teams where they don't apply, the platform is steadily becoming a more complete production runtime. The official cookbook now includes a road trip planner example demonstrating multi-agent coordination with the new features, and you can explore the full API interactively via the built-in claude-api skill in Claude Code.

Trending
  • No trending articles

Comments

avatar

Next Reads