Independent investing research and personal finance for IrelandUnderstand. Decide. Compound.

Wealth5 min read

Anthropic's Always-On Agent Blueprint: The Hidden Costs Are Permissions and Recovery

Anthropic's 13 November agent architecture guidance makes a subtle commercial point: subscriptions alone won't make unattended AI reliable without durable queues, permissions and recovery.

Anthropic chief executive Dario Amodei speaking at TechCrunch Disrupt in September 2023; archival photograph, not a demonstration of always-on agentsWealth / The Compound journal

Photo: TechCrunch / Wikimedia Commons; photograph by Kimberly White, CC BY 2.0 (https://creativecommons.org/licenses/by/2.0/); 20 September 2023, adapted crop · Illustrative image.

A system that can wake up each morning is easy to describe as an always-on AI agent. A system that never repeats the same payment, loses a user instruction or quietly misuses a credential is much harder to build. That gap is where the economics of enterprise agents may increasingly be decided.

On 13 November 2026, Anthropic published Building always-on agents, an engineering guide by William Hu. Its central message is architectural rather than a model launch: a proactive assistant needs a durable event inbox, timers, memory, permission controls and an outbox. Anthropic has not announced a separate new always-on agent revenue stream or guaranteed uptime as part of this post.

This differs from an app that answers a single question in one session. If the app is expected to monitor a queue, revisit unfinished work and occasionally ask permission, reliability must survive restarts, duplicate triggers and interrupted conversations.

The five pieces, and why each can become a cost centre

An event inbox records external triggers before invoking a model. If two webhook attempts arrive for the same invoice, the app should identify the duplicate rather than processing the payment twice. A timer system schedules follow-ups, but it must recover due work after downtime. Memory keeps durable instructions and task progress without endlessly replaying full chat histories.

Permissions constrain what the service can do when nobody is watching. The outbox records what should be sent to the user, and gives the product a way to recover when a notification service is temporarily unavailable.

All five introduce software and support obligations beyond purchasing model tokens. Someone must operate databases, logs, retries, credential handling, retention controls and incident response. For an enterprise buyer, this complexity is not a reason to reject agents; it is a reason to compare the entire workflow with the manual alternative.

Reliability means getting the same answer after a restart

A conventional chatbot error might mean regenerating one response. A proactive agent might already have sent an email, modified a CRM entry or created a support ticket before its session crashes. Retrying the whole task can duplicate an external action.

The most useful commercial question therefore becomes: can an organisation trace each event to a unique, authorised outcome? Products should support idempotency keys, explicit task states and a record of confirmed external actions. An inbox entry saying ‘processed’ is not proof that the intended destination received it; equally, a timeout does not prove an action failed.

Anthropic recommends a single active worker per inbox to avoid competing runs, and distinguishes the durable inbox from timers that merely make work eligible. Those patterns are helpful, but customers must still test how each connected third-party service behaves under partial failure.

The permission decision is a product feature

Giving an agent full account access may simplify a demo but increases the potential consequence of mistakes and prompt injection. A governed product needs scoped credentials, distinct limits for reading versus changing data, explicit user approvals for consequential actions and a reliable revocation path.

Anthropic’s guidance describes credential proxies outside the agent environment, automated risk screening and deferrable calls that resume when a user approves. These are proposed patterns, not a certification that every Anthropic-based agent automatically implements them.

The economic trade-off is real. A cautious approval flow may reduce the time saved. But an unattended service without strong approval boundaries can become more expensive through reversals, compliance issues and customer trust losses.

An illustrative unit-economics exercise

Suppose a customer automates 300 routine agent runs a day, over a 30-day month: 9,000 runs. At an assumed $0.04 in model/tool spending per run, the direct inference and tool bill is $360. This is a hypothetical price, not an Anthropic API quote.

Now assume 5% of runs require a human to review an exception. That is 450 escalations. At three minutes each, the organisation spends 22.5 staff-hours. Using an illustrative loaded labour cost of $40 per hour, review capacity is worth $900. Together those two components reach $1,260 a month, or 14 cents per attempted run, before databases, administration, software licences and incidents.

A vendor that highlights the four-cent model charge could materially understate real operational cost. Conversely, if the service replaces a larger and well-documented manual workload, the total economics could still be compelling. Buyers should compare cost per correctly completed, audited outcome, not cost per generated message.

A measurable way to pilot agents

Start with a reversible workflow such as collecting weekly project-status updates. Assign each task a durable identifier and specify how far the agent may go without permission. Run failure tests: duplicate an event, remove a token, simulate a timed-out connector, restart mid-task and send conflicting user instructions.

Track accepted outputs, duplicate side effects, missed jobs, correction burden, time waiting for approval and unsupported actions that the system successfully blocks. Also measure whether reports reach the correct users. A silent automation that completes internal steps but fails to notify the owner is a business failure even if its language output is excellent.

Anthropic’s documentation notes that an ‘always-on’ experience is actually built from short runs with idle periods. That matters to both cost forecasting and failure recovery. Charging customers for a persistent result while using event-driven execution could improve economics, provided reliable work is actually delivered.

The commercial opening and the constraint

The opportunity for Anthropic is deeper enterprise adoption: customers may use Claude-powered services not just to draft a document, but to maintain business processes across days. That can create recurring usage and implementation demand. Yet an architecture guide does not establish usage growth, pricing power, new contracts or profit margins.

Recent Claude service interruptions also show why production resilience deserves attention. This article asks a different question from our Claude Meetings analysis: what infrastructure and unit economics are necessary between conversations?

The key takeaway for enterprise teams is simple: an agent’s most expensive error is often not a wrong sentence, but an untracked action or a lost obligation. If proactive AI becomes a sustained business category, its best products will make these failure modes visible and manageable.

Explore the Anthropic company profile for broader private-company research.

Research checked 16 November 2026; source engineering guide published 13 November. Monetary scenarios are illustrative, not actual pricing or observed Anthropic customer results. Cover: TechCrunch/Kimberly White, September 2023, CC BY 2.0, adapted archival crop.

YOUR PRIVACY

Help us understand what gets read.

With your permission, anonymous Google Analytics measurements tell us which articles are useful. We do not send calculator entries or form values. You can decline and still read everything.