init: How Agentec approached AI agent deployment

When we introduced Agentec in August 2025, we described a design that separated the software running an AI agent from the settings for a particular workflow. A team could change a routing rule or revise instructions without rebuilding the software image that runs the agent. That separation also gave teams two places to look when a workflow’s behavior changed.

What the shared runtime was for

Our original design loaded each workflow’s configuration into the same runtime software. The configuration described the workflow’s instructions, model choice and available tools. The runtime handled execution. A configuration change could be tested against the workflow it affected. A runtime change, such as an update to how tools are invoked, called for checks across the workflows that used it.

The distinction is useful when troubleshooting. If a request goes to a different queue, compare the workflow configuration and the source the agent read. If both are the same, look at changes to the runtime or model and replay the request before rolling out further.

Test one rule change with a fictional request

Imagine a facilities team testing this request in a sandbox: “The loading entrance door won’t latch.” It arrives at 7:12 p.m. local time. The workflow can read a site guide and draft a maintenance ticket. A facilities coordinator approves any assignment.

1. Run the current configuration. Version 3 proposes the Facilities queue and cites the guide’s entry for door hardware.
2. Run the candidate with the same input. Version 4 adds an after-hours rule. It should propose On-call Facilities, show which rule it applied, and carry over the door and location details. Keep the model and source revision fixed so the routing change is easier to trace.
3. Test the approval boundary. The proposed assignment stays in review until the coordinator approves it. If the coordinator declines it, verify that the run records the decision and makes zero queue writes. Replay once with the site guide unavailable and check that the workflow asks a person to route the request.

Record the two configuration versions, the source revision, each proposal and the coordinator’s decision. That record gives the team a small regression test when the routing rules or runtime change again.

Apply the same test to your first workflow

Pick a repeatable request with a clear owner. Compare the expected result with the proposal and include one case that needs a person to step in. See the current Agentec platform overview, the review and approval controls, and the guide to planning an AI agent pilot.