One round of evolution
The system’s structure is rebuilt continuously, and every cycle consists of the same steps. Only steps linked as cause and effect run one after another. Everything else runs in parallel.
- A participant publishes an advertisement with a limited lifetime: what it accepts, what it produces, what it needs and how many more links it can take on.
- Every other participant independently scores the visible candidates by its own criteria.
- The pair negotiates in alternating moves: proposal, counter-proposal, agreement or refusal.
- A link on probation comes into being.
- The boundaries are synchronised in parallel with the system’s other links.
- The link’s strength and reliability are updated from the actual result.
- The parties review the terms: strengthen, weaken, put to sleep, dissolve.
Why topology is not assigned from above
The environment gives participants a way to meet but leaves the choice to them. The reason is practical. To choose the best structure centrally, you need to know the state of all participants at once. That is either expensive or inaccurate: while the state is being collected, it is already changing.
Local decisions do not produce the best structure. They produce a workable one that keeps adjusting. We assume that over the long run this is worth more than a one-off best structure for conditions that have already moved on. Only experiments can test that assumption.
What you see when you watch it
When you run such a system and watch it over time, some characteristic things show up:
- links gather around participants whose capabilities are in demand, and such hubs appear without any “coordinator” role;
- useless links do not break off at once but fade, so a temporary failure does not break the structure;
- after a participant is lost, the structure rebuilds itself within a few rounds;
- parallel branches of work do not get in each other’s way until they lay claim to a shared resource.
These are observations from our runs, not proven laws. The one thing we have measured rigorously so far argues against big conclusions: in a comparison with 16 agents and five seeds, the graph of links gave no advantage, 97.21% of tasks completed successfully against 97.24% without the graph, and the sign of the difference changed from seed to seed. The tasks there were single-agent, so whether a living graph helps where joint work is needed remains an open question.
Where a living graph is not needed
- A process with a fixed order of steps. It needs reproducibility; adaptation is not needed.
- A small number of participants. It is simpler to set the structure by hand and not pay for the negotiation machinery.
- Strict regulatory requirements. If the structure of interactions is set by regulation, it cannot be changed according to usefulness.
What this changes in how you think
The least familiar part here is giving up the question “which architecture is right”. In its place comes the question “which environment rules lead to the behaviour we need”. This is closer to designing an economy or the rules of a game than to drawing a flowchart: you set constraints and incentives, and the system finds the structure itself.
Hence the way to check it. The question is “what happened in the run”, not “does the system match the diagram”. A separate part of the series covers reproducible runs and evidence.