Agentic Systems

Under what structure can software agents execute multi-step business processes reliably, within limits that people define and can audit?

Business processes are long, stateful and full of exceptions. A single general-purpose agent is hard to test and harder to audit, and a process that depends on people triggering each step does not scale with volume.

Operations are naturally state machines: an item moves through states, and only some transitions are allowed. Autonomy becomes tractable when the permitted actions are explicit, each one is checked against policy before it runs, and anything outside the policy goes to a person.

A process is a finite state machine with states s, actions a and a transition function δ. A guard g encodes policy: it returns 1 only when the action is allowed in that state for the context x. Otherwise the item is escalated.

1. Decompose the process into states and allowed transitions.

2. Assign each transition to a specialized agent with a narrow role.

3. Express policy as guards and define what escalates.

4. Replay recorded event logs to test the guards before enabling autonomy.

1. Goal and policy

2. Planner splits the task

3. Specialized agents execute

4. Policy guard validates

5. Event log

6. Human escalation

Menlo Tech has not published experiments on this question. The Multi-Agent Coordination lab shows one computable piece of it, task allocation under capacity, on synthetic data.

Multi-Agent Coordination

No original research results are claimed. The only deployed evidence is the Subway collections operation, with published indicators (see the case study).

• A guard can only forbid what its author anticipated.
• An agent can still err inside permitted actions.
• Specification effort grows with the number of states.

Background reading by other authors. These are not Menlo Tech publications, and no Menlo Tech paper is cited because none has been published.

• Russell, S. & Norvig, P. Artificial Intelligence: A Modern Approach, 4th ed. Pearson, 2020.
• Wooldridge, M. An Introduction to MultiAgent Systems, 2nd ed. Wiley, 2009.