AI agents — recap & real-world examples

Real-world examples & war stories

Coding agents beating open-ended writing agents to reliability. Across the industry, agents that work inside a codebase — where a failing test gives immediate, unambiguous feedback — improved faster and became trustworthy sooner than agents built for open-ended writing or research tasks, where "did this work?" has no clean answer. 🎯 Takeaway: the environment's feedback quality predicts an agent's reliability better than the sophistication of its reasoning does.

A customer-service chatbot bound by a hallucinated policy. A well-known incident saw an airline's support chatbot invent a refund policy that didn't exist; a tribunal later held the company to the promise the bot had made up. 🎯 Takeaway: this is exactly what an unverifiable, high-stakes action with no human gate looks like in production — the fix isn't a smarter model, it's not letting an unsupervised agent make binding, irreversible commitments in the first place.

"Agent-washing" in vendor pitches. As agent hype grew, a recurring pattern of enterprise complaints emerged: products marketed as autonomous "agents" turned out, on inspection, to be fixed scripted workflows with a single model call inside them — or the reverse, genuinely open-ended systems marketed as predictable, scoped assistants. 🎯 Takeaway: placing a proposal honestly on the autonomy spectrum before signing off on it catches a mismatch between what was promised and what was built.

Early coding-agent benchmarks that looked strong on short tasks and collapsed on long ones. As agent benchmarks matured, a consistent pattern appeared: success rates that looked respectable on short, few-step tasks fell sharply as task length grew — the signature of compounding per-step error rather than any single dramatic failure. 🎯 Takeaway: demo-horizon thinking — judging a three-step demo as if it predicts thirty-step production performance — is one of the most consistent ways teams overestimate what an agent will do in the real world.

"Overfunded BPO" concerns in early agent startups. As service-as-a-software pitches proliferated, a recurring skeptical pattern emerged in due diligence: some AI services marketed as autonomous were, on closer inspection, staffed heavily by humans quietly doing the work behind the interface. 🎯 Takeaway: the honest supervised cost, not the marketed autonomy level, is the number that reveals whether an agent economy is real or subsidized.

Module recap

The through-line: the loop behind an agent is simple; almost everything that determines whether it succeeds is a decision made around it — how much autonomy it's given, how well it's set up to notice its own mistakes, and whether the task was ever a good economic fit for a loop in the first place. This module deliberately stayed at that decision altitude rather than re-deriving the mechanics already developed in full depth in Agentic AI for the AI PM, and rather than repeating what Tool calling and Memory & context already cover in this same family, because the mistakes that actually sink agent initiatives are rarely mechanical. They're a scope call made by demo appeal instead of economics, or a reliability bar set by a three-step test instead of the real thirty-step task.

Walk-away question: "For this agent proposal: have we placed it honestly on the autonomy spectrum, do we know what our real per-step reliability implies at its actual task length, and does the honest supervised cost beat what we're already doing today?"

If yes, this is an agent worth building. If no, you now know exactly which lesson in this module to reread — and where the deeper engineering lives, one module away in Agentic AI for the AI PM.