Unpopular Opinion- We are wasting immense engineering energy trying to make LLMs 100% deterministic.
by•
Every top-layer patch, Vector DBs, RAG, custom wrappers, tries to force a probabilistic model to act like a traditional API. Mathematically, it fails because:
Probability is non-deterministic by definition.
Models continuously drift as weights update and context accumulates.
Instead of pretending we can turn probabilistic models into predictable binaries, production AI needs to focus on Constrained Reliability through Auditability.
You need:
Causal Lineage: Querying why an agent took a specific branch (walking execution trees, not just reading raw logs).
State Snapshots: Isolating exact context windows to debug drift.
Human-in-the-Loop Interceptions: Triggering governance before rogue actions hit production.
21 views
Replies
I agree with the direction, but I wouldn’t throw out the reliability work entirely. The mistake is expecting the whole agent to behave like a deterministic API.
A better split is to let the model handle interpretation, planning, and drafting, while normal code controls permissions, schemas, budgets, idempotency, and which actions are actually allowed. Then evaluate behavior across many runs instead of expecting the exact same response every time.
Lineage and state snapshots are essential, but I’d add one more layer: every external action should have a clear risk level and a defined fallback. Low-risk actions can retry or recover automatically; anything irreversible should pause for approval. The goal probably isn’t identical outputs — it’s predictable boundaries and recoverable failures.
Exactly. The shift is from trying to eliminate uncertainty to engineering around it. LLMs are probabilistic systems, so reliability comes from making their behavior observable, traceable, and governable rather than forcing deterministic patterns on top.
The missing layer is often the runtime substrate: capturing agent state, preserving causal relationships between decisions and actions, and enabling interventions when confidence or policy boundaries are crossed.
This is where agent infrastructure needs to evolve from prompt orchestration toward systems built around execution lineage, state management, and control loops.