hebel. today · 4 November 202638 / 69
Career · Product Management Evolution
AI agents demand deterministic foundations over total rewrites
Engineering teams often mistake AI-native product development for a blank-slate opportunity, launching rip-and-replace rewrites that discard battle-tested backends for agent-first architectures. This approach frequently collapses when non-deterministic agents bypass business rules or introduce unacceptable…
From 4 newsletters · 1 min read · one thing to try
Engineering teams often mistake AI-native product development for a blank-slate opportunity, launching rip-and-replace rewrites that discard battle-tested backends for agent-first architectures. This approach frequently collapses when non-deterministic agents bypass business rules or introduce unacceptable latency (The MLnotes Newsletter). Dropbox avoids this by using an Intent Compiler Pattern, where the LLM handles fuzzy human language while the deterministic backend enforces strict business invariants, permissions, and transactional rollbacks through standard API calls.
“AI rarely creates new problems — it exposes the ones a team already had, like no automated testing or no clear process.”
Product Talk
Conversely, Salesforce advocates for an ADLC (Agent Development Lifecycle) that assumes continuous, iterative rebuilding of agents to counter model drift and evolving policy documentation (Product Growth). While Dropbox prioritizes architectural stability, Salesforce treats agentic behavior as a dynamic variable that requires constant observability and production-level correction. Both models reject the notion that AI replaces core product rigor, agreeing that automated testing and clear operational processes remain the bedrock of successful delivery (Product Talk).
The tension between these approaches highlights a broader consensus: AI exposes existing process failures rather than creating new ones. Discovery and customer-centric planning remain essential, as the speed of delivery does not reduce the need for product coherence or the ability to kill ideas quickly (TLDR Product).
- Salesforce uses 'Agent Script' to force deterministic outputs, ensuring that critical business metrics like pipeline totals remain consistent across all automated interactions (Product Growth).
- Figma manages complex AI-driven workflows using a collection of 31 skills built on Claude Code to automate context gathering and experiment generation (TLDR Product).
Try today · 30 min
Map your agent's current workflow to the Intent Compiler Pattern.
Identify one agent task that frequently breaks business rules and isolate the deterministic API call it should trigger instead of executing logic directly.
You need
- Current agent prompt or workflow logic
- List of existing backend API endpoints
- Defined business invariants (rules that must never be broken)
Steps
- Identify the specific business rule the agent currently violates.
- Define the required action as a strict, strongly typed schema.
- Restrict the agent to output only this schema rather than executing raw code.
- Route the output to your existing backend function for validation and execution.
- Verify that the backend rejects any request violating the invariant.