Agentic AI workflow design: why graph-based agents beat long prompts
Treating an agent as a long prompt gets you moving fast, but it's hard to trace what ran and why. London System Agent compiles a natural-language requirement into an explicit graph (IR) of input, model, tool, eval, and loop nodes.
Why a graph
A graph makes the execution path visible. Each node owns one responsibility and each edge reveals data flow, so you can point at the exact node where cost accrued or a step failed.
Graphs are also resilient to partial edits. Instead of rewriting the whole prompt, the refine path sends the current IR plus one new instruction and asks the model for a complete modified IR.
The beta is still honest about the boundary: chat refinement is not direct node editing in the UI. The returned IR must pass validation, and failed refinements preserve the existing graph while suggesting a safer alternative.
The compile gate
Structure checks run as part of compilation. Static properties — a single entry point, no unreachable nodes, loops with a termination condition — are verified before deployment.
This gate stops plausible-but-broken agents from slipping through to execution.
What comes next
Once the graph exists, simulation, evaluation, cost metering, and API execution build on the same IR. In beta, workflows with clear verification criteria — document processing and research agents — are the safest starting point.
Your turn to build
Compile your first agent graph from a single sentence.