AgenticArmy: A Multi-Agent Programming Workflow

Inspiration

Anyone who's used AI to program before knows it's pretty effective. So, it's only natural to wonder if you can create a whole dev team consisting solely for agents.

The problem is, there's no shared memory. Coordination is tough when all your agents are working alone.

Problem, solution. We can improve memory on a technical layer with Moorcheh's SDK, and create a structured workflow between agents that allows development while catching issues.

What it does

The workflow, in summary

Human provides goal; high-level process, success metrics, etc. Move on.
Planning agent runs; finds ways the goal can be implemented. Move on.
Prompt humans to approve/deny plans. If denied, return to step 2. If approved, move on.
Task coordinator agent runs; Splits up tasks among the agents to achieve the goal. Move on.
Conflict analysis agent runs, and tasks are assigned a conflict score based on how likely agents are to overwrite each other’s work. If a threshold is reached, return to step 4 to reassign agents (merge tasks?). If not, move on.
Coding agents run; tasks are completed in isolated environments. Move on.
Merge agent runs; merges individual agent commits. If it is too difficult, return to step 4. Else, move on.
QA agent runs; tests code for functionality. If unsuccessful, return to step 6. If successful, the process is finished until the next iteration.

Notable Features:

Conflict analysis agent anticipates coding agent disagreements
Human intervention only used to a) set the goal, and b) approve/deny proposed plans
A similarity score is calculated to determine the possibility of code overlap, allowing for reliable code merges