Workflow Bureaucratism

From the AI Research Compass wiki

Workflow Bureaucratism: LLM applications should follow explicit, predefined workflows and standard procedures, with every step, branch and handoff written down.

Beliefs

Workflow Bureaucratism trusts the flowchart. Every request is triaged, routed to the right department, processed according to a standard operating procedure, checked and filed. Branches are explicit, handoffs are logged, and nothing happens that is not in the diagram. Where agent enthusiasts give a model a goal and hope, the bureaucrat gives it a form to fill in. The result is slower to build and less exciting to demo, but it behaves the same way on Tuesday as it did on Monday.

It sits left of centre and at the very bottom, the Harness end: the steps are prompts and text, and the intelligence is mostly in the process design. Its Orthodox shade marks an established, mainstream school. Graph-based orchestration, SOP-driven multi-agent frameworks and declarative pipelines all belong here, and they remain the backbone of many production deployments precisely because auditors, compliance teams and on-call engineers like knowing what will happen next.

Arguments

  • Explicit workflows make behaviour predictable and auditable, which regulated and customer-facing deployments require.
  • Encoding standard operating procedures into multi-agent systems can reduce cascading errors compared with free-form agent conversation.
  • Declared pipelines can be optimized and evaluated step by step, so improvements are measurable rather than anecdotal.

Critiques

  • Rigid flowcharts handle only the cases their designers anticipated, and real inputs keep finding the gaps.
  • As models get better at planning, hard-coded workflows can cap performance below what a more autonomous agent achieves.
  • Maintaining large workflows becomes its own bureaucracy, with every model upgrade requiring a review of every step.

Representative works

Relationships

How to get it

Land on this cell in the full or pro test: x ≈ −0.40, y ≈ −1.00, paradigm ≈ −0.68.

A first draft. Corrections and better arguments are welcome through a pull request. Edit this entry on GitHub Edit its data