At Docflow, we think that buildtime agentic software systems are better than runtime agentic software systems for core business workflows.

Determinism. Traditional deterministic software orchestrates every step of every flow. By contrast, runtime agents may take different paths or perform different judgments notwithstanding identical inputs and context. Yet, business processes benefit from strict adherence to the correct business rules.

Artifacts. Buildtime agents produce runnable, auditable artifacts i.e. the code. Runtime agents, by contrast, take actions at runtime; their only artifacts are production logs, observability metrics and the actions themselves. Artifacts, simply put, explain the why i.e. Why did the workflow take path A vs path B? Was that decision particular to the data in some case or to the code representing the rules. Agentic runtime systems on the other hand are governed by matrix multiplications on numeric tensors. Why one token was produced vs another is difficult to ascertain to say the least, let alone why one tool call was taken over another.

Operational costs. All the talk of evals being hard all but goes away. Evaluation of traditional software systems has been perfected by the software community over decades. Version control, rolling releases, logging, observability, alerting, etc reduce the operational burden of maintaining reliable software systems.

Development cost. Unlike runtime agents, code must still get written, and there is an inherent cost to software development. The cost manifests in the time and resources building and maintaining software systems. While not completely ameliorating, buildtime agents help with both. Features that once took 6 weeks at Docflow are now taking 4 hours. Ongoing maintenance requests that used to take 1 day are taking 30 mins. We also have line of sight on a completely self-healing system.

Ultimately, the cost of (albeit, accelerated) development must be weighed against the operational and reliability cost of non-deterministic runtimes. 

At Docflow, the answer is quite simple. Our workflows underpin core business functions. Their business can only run if we do the right thing. Our customers cannot afford agents that make different judgment calls or take different paths when processing orders for medical equipment. If business rules can be encoded in traditional software, they should be.

That’s why we spend all of our infrastructure and system resources investing in buildtime AI. We make the building and maintenance of traditional software systems increasingly leveraged and streamlined on top of buildtime agents so that our customers get all the benefits of AI in terms of cost and power with none of the downsides.