Designing Agentic AI: Autonomy Needs Architecture
In the second of my blog posts looking at implementing agentic AI in the #intled sector, I wanted to explore how to design your agent workflow.
So once you've selected your candidate workflow according to the criteria discussed in the previous post, the real work starts. "Agentic" means the AI has some autonomy to work out how to get a job done rather than just following a fixed script. That's powerful and it's exactly why design matters so much before you build.
The MIT Implementing Agentic AI programme highlights five key areas (1) to consider:
Access: Which external systems does it actually need? A discovery/enquiry agent might touch a RAG-enabled knowledge base, a CRM, maybe calendar or comms tools.
Reasoning: Simple rules, a single LLM call or a full multi-step planning loop? More autonomy demands more reasoning capability and this can lead to interesting iterative process enhancement.
Memory: Context from the last message or the whole enquiry history? Probably somewhere in between (Goldilocks memory!)
Action: Read-only lookups are low-risk. Writing to a CRM or qualifying a lead isn't. Be deliberate about what's actually in scope.
Human in the loop: where does a person need to review or approve before something happens? Too little oversight and there’s potential risk. Too much and the agent's usefulness disappears.
The next critical decision is: buy, build, or stack? Off-the-shelf is fast but may need some tweaking to fit your objectives. Building gives control but costs (potentially a huge amount of!) time and expertise. Stacking (i.e. assembling existing tools and APIs) can be a solution but runs the risk of having to retrofit with a clunky outcome.
Right now the #intled sector is under pressure from major shifts in international student flows, challenging compliance environments and intense global competition. Finding the resources to build from scratch or working on a retrofit may not be possible or even desirable - and could potentially lead to suboptimal or underwhelming results.
As per the previous blog post in this series, given how fast AI capability is evolving, the best approach is probably to choose quick win, discrete processes and aim at rapid implementation with iterative improvements. This perhaps lends itself to buying off-the-shelf solutions so as to enhance your existing process workflows.
And one question that doesn't always get asked in #intled or perhaps any sector: how is this going to impact your teams? Getting that balance right in terms of how the new agentic workflow fits with your teams’ processes and the extent to which staff can really see and feel the benefits - and buy into the opportunities - is as much a design imperative as anything technical.
Involving your colleagues throughout the design, implementation and UAT stages is absolutely crucial - and can lead to innovative approaches that truly are greater than the sum or their parts.
In the next and final post in this series, I will look at how to scale your agentic AI workflows and ensure secure governance.
(1) MIT Sloan Implementing Agentic AI: Building Your Organizational Playbook program 2026-06-08
