Are AI agents deterministic?
AI agents are not deterministic. Deterministic means that the same input always gives the same result, as with a pocket calculator. An AI agent, a program that uses AI to carry out tasks, is probabilistic instead: it picks from several likely answers, so two runs of the same request can differ a little. You can reduce that, and you can keep it away from the steps where it would hurt.
Why does the same question get two answers?
Ask a colleague to write the same email on Monday and again on Friday. The two emails will say the same thing in different words. An AI behaves in a similar way. It does not look an answer up. It builds the answer word by word, and for each word it chooses one of several that fit.
A setting called temperature controls how freely it chooses. A high value gives livelier, more varied text. A low value keeps the AI close to the most likely wording.
With an agent the effect can grow, because an agent works in several steps. A small difference at the start changes what it reads next.
Does a low temperature solve it?
It makes the agent much more consistent. It does not guarantee identical results. And consistent is not the same as correct: an agent can give the same wrong answer every time. Consistency is still worth having, because a mistake that repeats is easy to find and fix.
Where variation is fine and where it is not
- Fine: the wording of emails and summaries, ideas, alternative phrasings.
- Fine if someone checks: sorting enquiries into categories, reading values out of a document.
- Not acceptable: prices, totals, tax, contract dates, who gets access, whether a payment goes out.
A worked example: one quote, split in two
A customer writes: "We'd like a quote for 40 licences. We have been with you for three years." If the agent does the whole task alone, the discount may differ from one day to the next. So split the task.
- Reading the email: the agent. It notes the product, the quantity and the customer.
- The discount: a fixed rule. A table says what 40 licences and three years are worth. It says the same tomorrow.
- The total: a calculation, never the AI.
- The covering email: the agent. Two different drafts are no problem here.
- Sending: a person reads the quote and confirms.
The customer gets a personal email and a price that anyone in the company can explain. The agent did the two steps that need language. Everything that must be the same every time was never left to chance.
Four ways to make an agent more consistent
- Lower the temperature for checking and sorting.
- Fix the form of the answer. "Choose one of these four categories" leaves little room to wander.
- Give examples for the borderline cases, so the agent decides them the same way each time.
- Keep the job small. One decision per agent varies less than five.
What this means for a small company
Do not ask whether an agent is deterministic. Ask which parts of the task must be. Put those into rules and calculations, and give the agent the reading and the writing. Then the variation is welcome where you want a human touch and absent where you need certainty.
How Worknano handles this
Worknano lets you keep the fixed part and the flexible part in one process. A workflow is a set of steps defined in advance: what starts it, under which conditions and what happens then. An agent is one step inside it. Each agent has its own temperature, lower for steady answers and higher for more variety. You can also give the agent examples to guide it on borderline cases.
- Workflows: triggers, conditions and actions
- Build your own agent: instructions, model and temperature
- AI agents and their output types
Create a free workspace and put a rule and an agent into the same workflow.