At some point, many agentic systems end up with a prompt that nobody wants to touch. What began as a few instructions gradually accumulates incident fixes, business rules, workflow logic, examples, exceptions, and model-specific caveats. Over time, the prompt becomes one of the most critical pieces of software in the system despite rarely being designed, reviewed, or maintained like software.

Drawing from experience operating production multi-agent systems, this talk examines how prompts evolve into accidental state machines, memory stores, configuration systems, and workflow engines. Through real examples, we’ll explore the failure modes this creates: degraded reliability, rising latency and cost, brittle behavior across model upgrades, and prompts that become impossible to reason about or safely modify.

The talk introduces a set of architectural patterns for keeping prompts focused on intent while deterministic systems own state, validation, business rules, and operational concerns. Rather than prompt tricks or model-specific techniques, the focus is on engineering decisions that determine whether an agent remains understandable and maintainable months after it ships.

Key Takeaways

Attendees will leave with practical heuristics for deciding what belongs in a prompt, what belongs in code, and how to prevent prompt-driven systems from turning into legacy software

Audience

AI/ML Engineers, Software Engineers, Platform Engineers, Architects, and Engineering Leaders building LLM-powered products, copilots, and multi-agent systems. The session is particularly relevant for teams moving beyond prototypes and discovering that maintaining agent behavior over time is often a harder problem than getting the first version to work.

Bio

Sujit Kamthe is a Solution Consultant at Sahaj Software based in Pune, India. He focuses on data engineering, large scale data processing, and architecting robust data platforms. He has previously presented at The Fifth Elephant on pragmatic guides to robust data quality checks, and his work centers on building enterprise systems that are scalable, governed, and ready for modern AI consumption.