When the executive who’s been most aggressive about agents tells Reuters that agent development is going slower than expected, that’s worth more than another demo. The gap between a single agent nailing a scripted task and a fleet of them running unattended in production is where most timelines quietly die. Here’s what actually eats the schedule, from the side that ships this:

• 🎯 Reliability compounds against you. A step at 95% success sounds fine until a ten-step chain lands you near 60%. Long-horizon autonomy is where the math turns brutal.
• 🔁 Recovery is harder than action. Getting an agent to call a tool is easy; getting it to notice it went wrong three steps ago and unwind cleanly is the real work.
• 📊 Evals lag the product. You can’t ship autonomy you can’t measure, and building the harness that catches silent failures takes longer than the agent.
• ⚠️ Guardrails aren’t free. PII, permissions, and rollback all add latency and edge cases that no keynote budgets for.
• ⚡ The last 10% is 90% of the calendar. Handling the weird inputs is the entire job once the happy path works.

None of this means agents don’t work. It means the demo-to-production tax is real and mostly invisible until you’re paying it. The HN discussion has the usual split between “told you so” and “still early,” and both can be right.

My bet: the teams who slip quietly on eval infrastructure now are the ones who ship reliable agents first. Who’s actually measuring the ten-step failure rate?