Two years after AutoGPT hype, here is what AI agents can really do in production — and what still fails systematically.
In 2023 AutoGPT was going to automate everything. In 2026, where do we stand?
Claude Code, Cursor Composer, Perplexity, browser agents — all specialized, not generalists.
Long-horizon tasks, human ambiguity, strategic decisions, silent errors.
Plan-then-execute, separate verifier, bounded loop, human in the loop on irreversible actions.
Three tools, three philosophies. I spent 3 months using them in parallel on real projects. Nuanced verdict with concrete use cases.
Going from 1 tool to 10 is not linear. Here are the patterns that prevent your agent from getting lost on the 5th call.
Reconciliation, reminders, VAT tracking, filings — how I automated 90% of my admin with a small homemade AI agent.
I design and ship tailor-made applications. 9 years of experience. Let's talk over a coffee (or a call).
Start a project