For teams that have tried deploying these in production environments, where do they actually break down?
Is it reliability over long action chains, tool integration issues, cost unpredictability, state management, latency, observability, or something else entirely?
I’m especially interested in failure modes that only became obvious after moving beyond controlled demos into real usage.
buschleague•1h ago
Constraint adherence degrades over long chains. You can put rules in system prompts, but agents follow them for the first few steps, then gradually drift. Instructions are suggestions. The longer the chain, the more they're ignored.
Cost unpredictability is real but solvable.
Ultimately, the systems need external enforcement rather than internal instruction. Markdown rules, or jinja templates etc., that the agent can read (and ignore) don't work at production scale. We ended up solving this by building Python enforcement gates that block task completion until acceptance criteria are verified, tests pass, and architecture limits are met. The core learning being that agents can't bypass what they don't control.