Agent Personality Engine for Multi-Agent Teams

Randy runs a small team of AI agents — Business, Technical, Personal — on top of his own modular agent infrastructure (MAR for execution, KGK for canonical evidence). Functionally they're distinct, but nothing gave them character: no persistent voice, no sense of which agent should notice or respond to a given event, and no principled way for the team's evolving beliefs about Randy to be trusted versus just asserted by whichever agent spoke first.

ape sits between that shared knowledge substrate and the execution layer as a decision layer, not a second runtime or a second memory system. Each agent gets a Character Profile — identity, voice, behavioral anchors, KGK-backed. An attention policy decides, per event, whether a given agent cares and how urgently; a deterministic arbitration step decides who actually holds the floor. Each agent then logs two agent-local facets that are never merged across agents: what it felt, and what it infers. Those inferences become trait candidates that only get promoted to a confirmed, shared belief when two or more agents corroborate independently, or Randy confirms directly — never a silent overwrite.

It's a real installable Python package (agent-personality-engine, Python 3.11+, MIT license), public at github.com/Yamai7354/ape. Ten-plus focused modules — character, attention, arbitration, composition, experience, corroboration, memory, audit, plus the extension seams for MAP (perception) and an external frontend — each with its own tests, and no storage of its own: every durable record is a KGK statement reached only through scope-enforcing gates.

The v1 core is complete and independently verified at 415 passing tests, including a dedicated suite that checks the architecture's boundaries hold end to end rather than just that individual functions behave.