Why AI Agents Still Need a Playbook

Originally published on my LinkedIn.

I was recently speaking with someone exploring how to automate parts of their procurement workflow using AI agents. Since I’m not deeply familiar with procurement, I asked how things are typically done today. They had already done their homework and had a workflow diagram for their current process handy!

Looking at it my first thought was that this is a fairly "solved" problem (at least if you're determined/desperate enough). After all, BPM tools, workflow engines, orchestration frameworks, and actor systems (like Erlang or the now-sunset Akka etc.) have been around for decades.

These traditional systems are deterministic by design. But also quite rigid. They rely on hard-coded logic and rule sets that are often difficult to update and can be limiting in complex or evolving environments. Their “degree of freedom” is essentially fixed at build time.

On the other end of the spectrum, we now have "free-form agents" powered by LLMs. These systems have a much wider (and fuzzier) range of behaviors. Their degree of freedom is defined by what the LLM can interpret and execute, now through dynamic tool use too.

In some sense, this is the classic bias-variance tradeoff. On one side, you have systems that are highly "biased" - predictable, constrained, and easy to reason about, but often inflexible. On the other, LLM-based agents can exhibit high variance - capable of novel and creative solutions, but harder to control or guarantee.

The key insight, I think, is that in real-world organizational settings, we need a thoughtful combination of both. Deterministic repeatability is as important, if not more, as creative flexibility. (Revealed preference: how many organizations prefer codified deterministic processes over creative execution by unarguably intelligent agents i.e. their own employees? In regulated industries?)

It’s still early days, but I believe the most successful agentic systems will be those that blend the strengths of both the approaches:

• grounded and deterministic guardrails to constrain the overall degrees-of-freedom of the system in critical, regulated, or safety-sensitive areas (while being flexible and underspecified in others)

• a collection of autonomous agents traversing these "safe" pathway to accomplish certain tasks and jobs with flexibility.