The author argues that autonomous AI agents are overrated without structured business context and scoped jobs, sharing practical insights from client work where agents run on fixed cadences with human oversight on writes.
I have been building around agents for client work for a while now, and my take is probably less exciting than the demo videos. I don't really want an agent waking up, looking around, and deciding what to do. At least not yet. That sounds cool until the work touches real accounts, client data, budgets, CRMs, tracking, websites, or anything where a bad write actually costs money. **The part I trust is structured context plus scoped jobs.** Every client has their own folder. Emails, meeting transcripts, call recordings, offer docs, pricing, website content, CRM notes, tracking notes, ad account data, conversion data, previous tests, all of it lives in one place. Most of it is pulled in automatically through n8n, Codex automations, or whatever connector makes sense for that client. The folder structure matters more than I expected. Same rough layout across clients, same naming conventions, same instruction files, same connection notes. When I open a client folder in Claude Code or Codex, the model is not starting from a blank chat. It can read the business first. **That makes the agent much less stupid.** It is not trying to reason from a prompt like "help this client grow." It can look at what the business is, what we tried before, what changed recently, what the CRM says, what the ad platforms say, what the last meeting was about, and then do a narrow job against that context. Stuff like: * daily account check * tracking audit * search term review * source health check * transcript into open actions * broken conversion handoff check * draft recommendations with evidence attached That is the part that compounds. If I improve the tracking audit once, I can run a better version of it across every client. If a weird edge case comes up in one account, it usually becomes a note or rule I can reuse somewhere else later. **I trust scheduled agents more than open-ended agents.** I tried the version where an agent wakes up, looks around, and decides what matters. It sounds cool. In practice I don't really trust it that much yet (give it 6 months tbh). Most of the useful stuff in my setup runs on a fixed cadence. Morning account checks. Weekly search term reviews. Monthly reporting passes. Tuesday and Thursday deeper account work. Some of it runs through Codex automations, some of it through n8n, some of it is still me manually kicking off the workflow. The agent is not the router. I am. The agent does the read work, runs the checks, drafts the output, and tells me what deserves attention. My alerts are mostly email and Telegram, not Slack. Daily account summaries go to my inbox. Telegram is useful when I want a quick pulse or to trigger something from my phone. If I need detail, I open the folder. **Writes stay gated.** Budget changes, paused campaigns, negative keywords, CRM writes, conversion settings, website deploys, anything that changes state or can cost the client money. The model can draft, stage, queue, explain. I still review before it goes live. That is not me being scared of automation. It is just the only version that survives contact with real accounts, platform policies, messy tracking, delayed conversion data, and clients who understandably do not want an agent freelancing inside their business. So I am less interested in "can the agent run 24/7?" and more interested in "does the agent have a structured place to work from, clear jobs, and hard approval gates?" Curious how others here are handling this. Are you building open-ended agents, or mostly scoped agents with structured memory/context underneath?
The author argues that in enterprise AI agent development, operational reliability and stability are more critical than high autonomy, advocating for controlled intelligence over fully autonomous systems.
After a year of building AI agents, the author argues that businesses don't actually want autonomous agents—they want specific outcomes like reduced support tickets and manual work, and simpler, reliable solutions often deliver more value than highly autonomous systems.
The article argues that AI agents need better judgment about when to refrain from acting, especially in contexts with incomplete data or irreversible outcomes, and that controlled autonomy is more trustworthy for companies.
The article questions why most autonomous agents are developed for business use rather than for individual users, pointing out a gap in AI accessibility.
The author reflects on the current limitations of AI agents for complex, long-running tasks, citing reliability issues and suggesting that agents are better suited for narrow, supervised tasks rather than full autonomy.