Blog
Guides, best practices, and product updates from Anlyon.
What happens when an AI agent retries a payment that already succeeded
An agent requests a refund, Stripe executes it, and the response times out. The agent retries. Here is why an idempotency key alone does not save you, what Anlyon records instead of guessing, how an unknown outcome settles, and where the design stops.
What happens when an AI agent's API call times out
A timed-out refund may already have happened. Why retrying an agent's side effects is dangerous, failed versus unknown, and how idempotency keys make it safe.
An AI agent kill switch: what it has to stop, and what it can't
How to build a kill switch that stops an AI agent in production, including queued work and approved calls. What a real halt must reach and what it can't recall.
AI agent staging and production isolation: why the key must pick the environment
Keep an AI agent's staging key away from production data and side effects: bind the environment to the credential, remove every fallback, and test the boundary.
How to add human-in-the-loop approval to AI agent tool calls
Pause an AI agent's tool call until a human approves, with the agent unable to approve itself and every decision recorded. Two ways to build it, in TypeScript.
How to let an AI agent call the Stripe API without giving it your key
Let your AI agent issue Stripe refunds without holding the secret key. The agent names an action. Anlyon resolves the credential and makes the call.
EU AI Act Article 14 for engineers: what "human oversight" means in code
EU AI Act high-risk obligations apply from 2 December 2027. Articles 12 and 14 mean logging and a human who can override and halt, mapped to what you build.
An agent deleted a production database in nine seconds
The PocketOS incident had three missing guardrails. What each is, why the system prompt didn't help, and how to add them before your agent finds the gap.
We shipped an idempotency bug that ran your side effects twice
Our idempotency layer had a check-then-act race, in a product built to not do things twice. The bug, the fix, the test that proves it, and what else we found.
Your agent asks. Anlyon makes the call.
Models reason well, remember nothing, hold no credentials and don't know when to stop. Your agent names an action. Anlyon holds the key and makes the call.
Introducing Memory: remember() and recall() for AI agents
Long-term memory for AI agents in two function calls, remember() and recall(). No vector database, no embedding pipeline, no index tuning.
Knowing what your agent actually did
Delivery rates, failures and usage for work Anlyon does for your agent, plus a record of every production call it made: the operational half of an audit trail.
Where your agent's failed work goes
Messages, schedules and workflow steps that exhaust their retries land in a dead-letter queue you can inspect and replay. Why action invocations never do.
One agent decision, many downstream systems
Publish once and deliver to every endpoint that cares, each with its own retries and delivery tracking. For agents that have to tell more than one system.
Give your agent a schedule it can't forget
Nightly reports, hourly checks, follow-ups: recurring agent work without running a scheduler. Cron expressions, retries, and delivery you can audit.
The TypeScript SDK for agents that do things
One typed TypeScript client for letting an AI agent act on production: invoke an action by name, Anlyon holds the credential, and risky calls wait for a human.
Your agent's first side effect: a Stripe refund without the Stripe key
Let an AI agent issue a real refund without holding the credential. Define the call once, the agent invokes it by name, and Anlyon makes it after a human approves.