RSSAmplifier

Blog

Muness Castle

Muness Castle builds AI systems with the people who do the work, develops public tools, and writes about what he learns while building them.

muness.comRSS feed ↗5 posts

Latest posts

When AI Changes the Work, the Complements Change Too

In July 2025, Commonwealth Bank of Australia (CBA) told 45 customer-service employees that an AI voice bot had made their jobs redundant. Within weeks, the bank reversed the decision. It admitted that it had not considered all the relevant business needs and had not been thorough enough in deciding which roles were still required. Workers told their union that calls had increased while the bank...

AI: Token or GPT?

In April 2025, Duolingo CEO Luis von Ahn wrote to employees: “Duolingo is going to be AI-first.” The same memo said Duolingo would “gradually stop using contractors to do work that AI can handle” and grant headcount only when a team could not “automate more of their work.” Shopify described Tobi Lütke’s internal memo with a similar rule: “reflexive AI usage is now a baseline expectation at Shop...

Agent Work That Changes Behavior

The three posts in this series are trying to name one operating model for agent work: preserve the parts of learning that should survive a run, then make the next run behave differently because of them. Drazen’s The Context Stack names the underlying bottleneck cleanly: Action is cheap. Knowing what to do is scarce. That still feels like the core shift. LLMs made a lot of action cheaper:...

There Is No Agent Workflow Runtime

People keep saying “agent runtime” when the thing they need is a workflow runtime. That sounds like a naming complaint. I do not think it is. Names carry architecture. If we call the runtime an agent runtime, we tend to put the agent at the center of the system and make everything else orbit around it: tools, memory, prompts, approval, logs, tests, state, and deployment. Then deterministic wor...

Agents Don’t Learn the Domain. The System Does.

A data product is already a kind of memory. Not memory in the LLM sense. Memory in the operational sense: records arriving from several source systems, ingest paths, normalization routines, cleanup rules, custom pipeline steps, exception lists, review decisions, and the people who know why an apparently wrong adjustment is actually right. The business outcome is not the pipeline. It is a trus...