AI agents that run the same way every time.
Routine steps, like creating a sales order, run as fixed actions. AI handles the parts that change from run to run, like reading a PDF. If a step fails, the agent retries it or asks someone.
Example: entering a PO from an emailed PDF
Two kinds of steps in every workflow.
When you build an agent, Trelium separates the steps it can do the same way every time from the ones that need a model.
Run the same way every time
Steps with a known input and output, like creating a record or calling an API. They behave the same on every run and don't call a model.
- Look up a customer in NetSuite
- Create a sales order
- Update a status field
- Send a templated reply
Handle input that varies
Steps where the input changes from run to run, like a PDF from a new customer or an item code that isn't in your system yet.
- Pull order lines from a PDF
- Match an unfamiliar item code
- Classify an inbound request
- Draft a reply that needs context
Predictable runs
The agent doesn't plan the job again on each run. It follows the steps you reviewed.
Credits go to AI steps
Routine steps don't call a model, so you aren't paying model costs for data entry.
Readable steps
You can read the full list of steps before the agent runs and edit any of them.
How this compares to other approaches.
Chat-style agents can read unstructured input but plan each run from scratch. Trigger-based tools run the same steps every time but fail on input they weren't set up for. Trelium uses fixed steps and AI steps in the same workflow.
See detailed comparisons
| Capability | Chat-style agent | Trigger-based automation | Trelium |
|---|---|---|---|
| Messy inputs like emails and PDFs | Handles them | Breaks on anything unexpected | AI steps handle them |
| Same steps on every run | Re-plans each time | Yes | Yes |
| Cost per run | Model on every step | Low | Model only on judgment steps |
| When a system changes or fails | Varies | Stops with an error | Retries, then asks a person |
| Review before it runs | No fixed steps to review | Yes | Yes, step by step |
What happens when a step fails.
A system times out, a page layout changes, or a value comes back in the wrong format. The agent checks the error, retries if the fix is clear, and flags the run for a person if it isn't.
Run log
PO Entry Agent · #4821- 9:42:08Create sales order failedNetSuite request timed outFailed
- 9:42:09Checked the errorTemporary timeout, order not createdInvestigating
- 9:42:40Retried the stepSO-20391 createdRecovered
- 9:42:41Continued the runConfirmation sent to customerDone
- 01
Check
The agent reads the error to tell a temporary problem, like a timeout, from one that needs a change.
- 02
Fix
If the cause is clear, it adjusts the step, for example by finding a field that moved on a page.
- 03
Retry
It runs the step again and continues with the rest of the workflow.
- 04
Escalate
If the step still fails, the run stops and is flagged for someone, with a note on what the agent tried.
Review the steps before it goes live.
Each agent's steps are written out, so you can check them the way you'd check any written process.
See which steps use AI
Each step shows the system it touches and whether it calls a model. Edit any step before the first run.
Credits per run
Each run lists the tasks the agent completed and the credits it used.
Approval steps
Add a human approval before a step that sends an email, posts a record, or makes a payment. The run waits until someone approves.
Edit one step at a time
When your process changes, update the step that changed. The other steps stay as they are.
FAQ
Common questions about deterministic agents
It's an agent that follows a fixed, readable set of steps for work it already knows how to do, instead of working out the job from scratch every run. In Trelium, known steps run as software actions and the model is only used for steps that need judgment.
Yes. Reading emails and PDFs, matching values that don't line up, classifying requests, and drafting replies all use AI. Entering data and updating records don't.
Most steps in a business workflow are routine. Those run as fixed actions and don't use model credits, so credits are only spent on the AI steps.
The agent checks the error, fixes the step if the cause is clear, and retries it. If the step still fails, the run stops and is flagged for someone to look at.
Yes. The workflow is laid out step by step, with the system each step touches and whether it uses AI. You can edit it before it goes live and change it later.
Trigger-based tools run fixed steps but need structured input. Trelium uses fixed steps too, and adds AI steps for input like emails and PDFs. Failed steps are retried, and you can add human approvals in the same workflow.
Have a workflow your team runs every week?Bring it to the demo.
We'll go through it step by step and show which parts would run as fixed actions and which would use AI.