Quick answer#
Claude Managed Agents is a strong candidate for teams that want Anthropic to operate the Claude agent harness, while retaining application-level control over tasks and tools. It is not a Claude Pro subscription, a training course, or simply the Agent SDK deployed for you. The current beta provides agents, environments, sessions and events, with either Anthropic cloud sandboxes or self-hosted tool execution.[7]
Verified October 6, 2026. Pricing is standard model tokens plus $0.08 per running session-hour, with web searches charged separately. The headline runtime fee is not the complete task price. More importantly, current documentation says fast-mode premiums and US inference-geography pricing do apply; older comparison articles saying neither applies are no longer a safe budgeting source.[8][24]
- Best for: Claude-first development teams building long-running, asynchronous workflows without maintaining the entire agent loop.[7]
- Avoid if: you need zero data retention or HIPAA BAA coverage for this product, fully offline execution, non-Claude model portability, or a settled non-beta contract.[7][9][17]
- Main advantage: a managed harness, persistent session history, built-in tools and a documented session spend-control mechanism.[7][22]
- Main drawback: infrastructure abstraction does not eliminate model/token costs, tool authorization or provider-side data flows.[8][9]
This is a source-based review. We did not purchase a plan, deploy a session, benchmark reliability or measure a refund outcome. For direct model access and a loop you operate, the separate Claude for AI Agents review owns that API/SDK decision; this page evaluates the managed service.
What you are buying, and who operates it#
Anthropic contrasts the Messages API's direct prompting access with Managed Agents' prebuilt configurable harness. An agent defines its model, prompt, tools, MCP servers and skills; an environment defines where execution happens; a session performs work; events carry messages, results and status. The platform persists event history and supports steering or interruption.[7]
The default cloud environment includes bash, file operations, web search/fetch and MCP integration. The current overview also documents scheduled deployments. Those are product capabilities, not evidence that any proposed business process works reliably without review.[7]
The model remains Claude: agent setup supports Claude 4.5 and later models, with model-specific configuration and availability to check. Self-hosted sandboxes move tool execution, not inference or orchestration, onto your infrastructure. This is not an open-weight model download or an offline local agent.[31][9]
Anthropic's commercial terms identify the contracting entity as Anthropic Ireland, Limited for customers in the EEA, Switzerland or UK, and Anthropic, PBC elsewhere. Consumer Claude offerings have separate terms. A familiar company name is useful identity evidence, not a guarantee that a beta is appropriate for your regulated workload.[11][17]
Pricing: tokens, runtime and the modifiers older reviews miss#
The official pricing page bills every session's model tokens at listed rates, with caching multipliers, plus running-state duration at $0.08 per session-hour. Runtime is measured to the millisecond. Time in idle, rescheduling or terminated does not count; runtime replaces the separate code-execution container-hour charge inside Managed Agents.[8]
Current representative USD model rates per million tokens are:[8]
- Claude Sonnet 5.5: $2 input, $10 output, $2.50 five-minute cache writes and $0.20 cache hits/refreshes.[8]
- Claude Opus 5.5: $4 input, $20 output, $5 five-minute cache writes and $0.20 cache hits/refreshes.[8]
- Claude Haiku 4.5: $1 input, $5 output, $1.25 five-minute cache writes and $0.10 cache hits/refreshes.[8]
These are model-price references, not a promise that every account has access to every model or that every request has the same modifiers. The model configuration and actual usage fields must determine the estimate.[31][8]
The important current exceptions are specific. The Batch API discount does not apply to these stateful interactive sessions. But fast-mode premium pricing does apply when model.speed is "fast", and model.inference_geo: "us" applies a 1.1x token multiplier. The same page distinguishes Anthropic-operated Claude Platform on AWS billing from partner-operated cloud platforms; Managed Agents token/runtime charges on Claude Platform on AWS convert into Claude Consumption Units.[8]
That distinction prevents two expensive mistakes: assuming a Claude subscription includes Managed Agents API use, or transferring ordinary Bedrock model pricing into this product. Claude Platform on AWS is not Amazon Bedrock Managed Agents powered by OpenAI. The latter is a different vendor-operated harness and billing surface.[7][8][21]
Hidden costs: a worked budget, not an average-user bill#
Assume a one-hour Sonnet 5.5 session with 50,000 standard input tokens and 15,000 output tokens, no caching, no fast mode, global inference and no web searches. Arithmetic on the published rates gives $0.10 input + $0.15 output + $0.08 runtime = $0.33. This is an illustrative workload, not our observed session consumption.[8]
With only the US inference modifier changed, that example becomes $0.355 before tax or other charges: the token subtotal is multiplied by 1.1, not the session runtime. Real runs can read and regenerate far more tokens through repeated tools, retries and growing context. Do not present the example as a normal cost per completed task.[8]
Other budget lines matter:
- Web search: $10 per 1,000 searches, plus the relevant tokens. Search-heavy workflows need their own usage allowance.[8]
- Cache behavior: cache writes and cache hits have different rates, and current model-specific cache prices differ. “Everything cached is 90% cheaper” is not a safe universal formula.[8]
- Concurrent activity: session-level active seconds count overlapping threads once; per-thread usage excludes session running-time cost. Do not sum rounded thread totals and expect an exact invoice reconciliation.[22]
- Your infrastructure: self-hosted workers, networking and operations remain yours. A vendor session charge is not an all-inclusive quote for the machine you run.[9]
- External tools and review: connected services may bill independently, and checking a wrong or unsafe output is still work. Anthropic's terms leave output suitability and appropriate human review with the customer.[11]
One hundred running session-hours alone calculate to $8, before tokens and tools. Use the agent cost worksheet to track total spend per accepted output rather than comparing that runtime subtotal with a complete hosting bill.
Session budgets: useful protection with an overshoot boundary#
The current budget API is materially more useful than an advisory “try to spend less” prompt. Set a budget when creating the session; its max_list_cost.amount is a positive whole number of US cents written as a string. The platform tracks model tokens, web searches and session running time at public list prices.[22]
But the stop is between model requests, not mid-request. An admitted request finishes, so spend can exceed the threshold by up to one request per thread. Reaching the cap idles the session with budget_reached; it does not delete the transcript or sandbox. The docs give the example of a 50-cent cap settling at 53 cents.[22]
Two lifecycle rules deserve an implementation test:
- You cannot add a budget to a session created without one. Attach it at creation, not after noticing runaway spend.[22]
- Removing a budget is one-way for that session: it cannot be re-added. Changing the cap is the safer operation if you intend to retain a limit.[22]
Negotiated discounts do not change the list-price basis of this control. Treat it as a limit on admitting new work, not a perfectly exact invoice ceiling or a cap on third-party systems. For monitoring, distinguish a clean budget pause from a failed job or an unreachable worker; the agent monitoring guide provides the broader acceptance and alerting framework.
Security and retention: self-hosted tools are not zero retention#
The overview explicitly says Managed Agents is currently not eligible for Zero Data Retention or HIPAA BAA coverage, because it stores state, history and outputs server-side. Sessions can be deleted, and separately uploaded files have their own deletion operation. Do not replace those product-specific statements with a generic claim that the Claude API supports a particular agreement.[7]
For self-hosted execution, files, processes and network traffic stay in your environment, but tool inputs and outputs still flow to Anthropic's control plane so the model can decide what to do next. Anthropic also stores skills and memory stores, with a copy downloaded into the sandbox. The boundary is therefore more nuanced than “my data never leaves my server.”[9]
The worker documentation says to keep the general Claude API key off the worker host, where agent tool calls could read it, and use an environment key for the worker. It also separates your network policy and lifecycle from Anthropic-managed cloud execution. Those are controls you must implement, not automatic benefits of the word “self-hosted.”[9]
Use the MCP security checklist to review credentials, scopes and destructive actions. Anthropic's reference supports tool confirmations and remote MCP connections, but a confirmation mechanism is only useful if your application enforces it and a real owner reviews consequential requests.[16]
Availability, limits and what beta means commercially#
The beta is enabled by default for API accounts and requires managed-agents-2026-04-01 on requests. The overview identifies MCP tunnels and dreaming as more limited research previews. Do not assume the main beta's accessibility extends to every advertised capability.[7]
Current reference limits are 300 create requests per minute and 1,200 read requests per minute, per organization, with organization spend limits and usage-tier rate limits also applying. Those are endpoint request limits, not guaranteed concurrent session counts or task throughput.[16]
Anthropic's service-specific beta terms say these services may be unsuitable for production, are temporary and provided as-is, and carry no obligation to provide indemnity. They cap beta-related liability at the lesser of $1,000 and the fees paid in the preceding 12 months. Have the responsible buyer review that contract boundary before using the product for a material customer obligation.[17]
Cancellation, prepaid credits and leaving the service#
The commercial agreement allows customer termination with notice, but incurred fees remain a separate obligation. If you use prepaid credits, the published credit terms describe them as non-refundable, expiring one calendar year after issuance for usage credits, and expiring on account closure. Auto-reload can charge the configured amount until you opt out.[11][12]
Do not apply prepaid-credit rules blindly to every billing arrangement. The current pricing page describes Claude Platform on AWS as arrears-only marketplace billing without prepaid credits. Confirm the actual account, offer and contracting terms before treating a cloud invoice and a direct API balance as the same thing.[8]
Before an exit, save required outputs, stop new work, review active sessions and any deployment triggers, delete stored sessions and uploaded files as appropriate, and inspect your own workers and external services. Those are proposed operational steps; this review did not cancel an account or test a refund. Merely abandoning a chat tab is not a documented shutdown process.
Independent evidence: distinguish concerns from measured defects#
In the launch-era Hacker News discussion, commenter pdp described the service as infrastructure for companies embedding agents and identified the Anthropic-model dependency as a trade-off. That is a useful architecture question, not a measured reliability result. The separate overview thread debates harness quality and subscription lock-in, with disagreement rather than a representative satisfaction sample.[13][14]
We did not find enough product-specific, independently reproduced runtime evidence in this pass to report a failure rate or typical cost. Targeted Reddit search returned no results; that limits coverage and does not prove absence of users or problems. Claims about Claude Code bugs or consumer-plan limits are not automatically Managed Agents complaints.
A more concrete evidence check is pricing drift. TrueFoundry's comparison says fast-mode and data-residency modifiers do not carry over and includes an older Sonnet 5 increase schedule. The current official page explicitly applies those modifiers and lists Sonnet 5 at $2/$10. We use the current primary source, not the competitor's stale rate table. TrueFoundry sells an alternative, so its comparison is commercially interested analysis, not neutral customer telemetry.[24][8]
Advertising evidence: a brand ad is not a Managed Agents result#
The Google Ads Transparency lookup for anthropic.com displayed a total of five ads and verified Anthropic, PBC cards. We captured the result page, not five individually inspected creatives. This is brand-level advertising evidence only; it does not establish a Managed Agents campaign or substantiate a performance claim.[27]
The Meta keyword query returned HTTP 403 while rendering a broad result list including unrelated advertisers. We did not classify those ads as Anthropic's or publish their library IDs as product evidence. Exact-offer paid-ad activity remains unverified, and we make no claim about spend, targeting or return on ad spend.[29]
Pros, cons and the strongest alternatives#
Pros: an integrated Claude harness and tools, stateful asynchronous sessions, choice of cloud or self-hosted execution, and a documented budget pause. These reduce pieces of infrastructure you would otherwise need to design.[7][9][22]
Cons: Claude dependency, beta contract risk, provider-side state that excludes current ZDR/BAA eligibility, and a complete bill that still includes tokens, modifiers and connected systems.[7][8][17] You also retain the business-level responsibility for authorization and correctness.[11]
- For AWS-native OpenAI work, read the Amazon Bedrock Managed Agents review. Its preview restrictions, IAM roles and infrastructure costs are different; it is not this service under another name.
- For custom code hosted within an Azure operating model, the Foundry Agent Service review compares the session compute and identity boundary. Compare complete responsibilities rather than unlike hourly units.
- For an operator-facing agent you can run yourself, use the Hermes Agent review. Hermes's official documentation describes its skill-learning loop and local installation; it is a different application choice, not a compatible replacement for Anthropic's managed-session API.[30]
If you want to operate that alternative yourself, follow the Hermes installation guide before evaluating a hosting purchase. Managed Hermes hosting is relevant when maintaining your agent server is the bottleneck. It does not automatically solve application development, regulated-data eligibility or compatibility with an existing Managed Agents integration. Conversely, if your team already wants Claude-native embedded sessions, there is no need to force an unrelated migration.
Decision checklist: make one pilot earn expansion#
- Define one accepted output and the human decisions that the agent must not make alone.
- Confirm the model, geography, fast-mode setting, account billing path and beta feature access before estimating costs.
- Create the session with a budget; allow margin for an in-flight request per thread and external-tool charges.
- Test approval denial, interruption, budget pause and resumption using non-sensitive fixtures.
- Verify what data is sent to Anthropic and what remains in your sandbox. Stop if the current ZDR/BAA exclusion conflicts with your requirements.
- Measure token usage, running time, external-tool charges, reviewer time and accepted outputs together.
- Test deletion/export and a controlled exit, then decide whether saved engineering effort justifies the provider dependency.
Verdict: worth piloting for a Claude-first team with clear acceptance criteria and compatible data requirements. Avoid choosing it solely because $0.08 sounds cheap, or rejecting it because an older article describes limitations the current documentation has changed.
Sources#
[7] https://platform.claude.com/docs/en/managed-agents/overview — Claude Managed Agents overview - Claude Platform Docs
[8] https://platform.claude.com/docs/en/about-claude/pricing — Pricing - Claude Platform Docs
[9] https://platform.claude.com/docs/en/managed-agents/self-hosted-sandboxes — Self-hosted sandboxes - Claude Platform Docs
[11] https://www.anthropic.com/legal/commercial-terms — Commercial Terms of Service \ Anthropic
[12] https://www.anthropic.com/legal/credit-terms — Supplemental Credit Terms \ Anthropic
[13] https://news.ycombinator.com/item?id=47693047 — Claude Managed Agents | Hacker News
[14] https://news.ycombinator.com/item?id=47697641 — Claude Managed Agents Overview | Hacker News
[16] https://platform.claude.com/docs/en/managed-agents/reference — Reference - Claude Platform Docs
[17] https://www.anthropic.com/legal/service-specific-terms — Service Specific Terms \ Anthropic
[21] https://aws.amazon.com/about-aws/whats-new/2026/09/bedrock-managed-agents-preview
[22] https://platform.claude.com/docs/en/managed-agents/budgets
[24] https://www.truefoundry.com/blog/claude-managed-agents-pricing
[27] https://adstransparency.google.com/?region=anywhere&domain=anthropic.com&hl=en — anthropic-google advertising lookup
[29] https://www.facebook.com/ads/library/?active_status=active&ad_type=all&country=ALL&q=Claude%20Managed%20Agents&search_type=keyword_unordered — claude-meta advertising lookup
[30] https://hermes-agent.nousresearch.com/docs — Official Hermes documentation
[31] https://platform.claude.com/docs/en/managed-agents/agent-setup — Claude Managed Agents model and agent setup