Skip to main content
Determinism is: same input, same output. LLMs are probabilistic, so you’ll never get bit-for-bit determinism from the model itself — and that’s fine. The goal isn’t a deterministic model, it’s a repeatable system: given the same situation, the agent reliably does the same reasonable thing. Most of the variance you actually feel comes from around the model, and that part you can control.

​
Where non-determinism actually comes from

If the agent sees different context on different runs, it behaves differently — even at temperature 0. Fuzzy, wide, or ranked-by-vibes retrieval is the largest source of “why did it do something different this time?”
When the harness lets the model decide everything — including steps that should be fixed — you get a different path each run. Some decisions belong in code, not in the prompt.
Ungrounded reasoning fills gaps by guessing, and guesses vary. Grounded reasoning over retrieved facts converges.
Temperature and sampling. Real, but usually the smallest contributor — and the one people reach for first while ignoring the larger three.

​
Push variance out of the loop

1

Make retrieval stable and precise

Deterministic ranking (explicit field boosts, not opaque relevance), narrow result sets, and addressed records mean the agent tends to see the same context for the same query. Stable inputs → stable behavior. → Context search accuracy
2

Fix what should be fixed in the harness

Encode the parts of the workflow that are genuinely fixed as control flow, and let the model decide only the genuinely open parts. Determinism where you can afford it buys you reliability where you can’t. → The harness
3

Ground every decision in retrieved facts

An agent reasoning over specific, cited entities converges far more than one improvising from a vague prompt. Grounding is a determinism lever, not just an accuracy one. → Provenance
4

Verify with a critic

A critic sub-agent that rejects outputs unsupported by context collapses the tail of weird one-off behaviors into a consistent “checked” band.

​
A useful reframing

You don’t make the model deterministic. You shrink the space it’s allowed to be non-deterministic in — by stabilizing its inputs (retrieval), fixing its scaffolding (harness), and grounding its reasoning (context).
The more of the surrounding system is structured and repeatable, the more the model’s residual randomness stops mattering.

Next: Self-learning loops

Agents that improve with use without drifting.