Agents need control flow, not more prompts
If you’ve typed MANDATORY or DO NOT SKIP into a prompt, you’ve hit the ceiling. Reliable agents need deterministic scaffolds encoding logic in software, treating LLMs as components, not systems.
What Matters
- Prompt chains are non-deterministic, weakly specified, and unverifiable; complexity collapses reliability without explicit state transitions and validation checkpoints.
- Without programmatic verification, your only options are: keep a human babysitter in the loop, run exhaustive post-hoc audits, or vibe-accept the outputs.
- LLMs as runtime orchestrators vs. LLMs as code generators is a key architectural fork: push hard business logic into generated, testable software.
- [HN: @827a] A QA agent iterating 200 markdown requirement files failed repeated prompt-based enforcement; only deterministic flow fixed it.
- [HN: @gck1] Counterpoint from a practitioner who went full circle: separate enforcement agents can handle DO NOT SKIP failures without abandoning prompting entirely—just decompose agent responsibility.
- [HN: @scrappyjoe] Week 1 of agent integration was expanding prompts and degrading performance; week 2 was defining precise objects and methods—mirroring the source thesis empirically.
- Google ADK v2 reorients around graph operators for control flow and includes parallel fan-out/join patterns directly in its docs.
