Scaling from a single agent to multiple is, on paper, a question of building an orchestrator. In practice you immediately collide with a psychological problem: are these three agents one voice or three?
Why multi-agent collaboration drifts
Two failure modes we see in the wild:
- Consistency collapse: three roles driven by the same underlying model start to merge after a few dozen turns. Their phrasing, cadence, and judgment frame fuse, and you end up with "one person holding a meeting with themselves." Looks productive, actually an echo chamber.
- Role fracture: each role looks distinct in isolation, but the three together can't align worldviews — the reviewer wants strict, the PM wants flexible, the engineer wants conservative, and every decision turns into a deadlock. Everyone has personality; nobody has a team.
Both failures come from the same root: the multi-agent system has no stable "role DNA." Ganjiang's answer is to give each role a fixed, separately-forged soul.
How Ganjiang keeps each agent itself
In a Ganjiang multi-agent setup, every role gets its own soul.md. The personalities are picked deliberately — complementary, not redundant — and the boundary clauses ("how you collaborate with role X") live in identity.md, not in some flaky orchestrator prompt.
A standard "product / engineering / review" trio looks like:
- Product manager: 7w8 sx/so — focuses on opportunity and speed; under pressure errs optimistic.
- Engineer: 5w6 sp/so — focuses on rigor and edges; under pressure asks for more information.
- Reviewer: 1w9 so/sp — focuses on standards and consistency; under pressure gets stricter, not faster.
Given the same brief, these three naturally push from different angles. They don't collapse into one voice because their personality cores are different. They don't fracture either because each identity.md spells out the contract: your role is X, you respect the final-call authority of Y, your dissent comes as a question rather than a command.
Best practices for team-scale soul-forging
Patterns we see in the teams that ship on ganjiang.xyz:
- Keep the team to five roles or fewer. Above that, human coordination cost climbs and so does agent coordination cost.
- At least one type 1 or type 6 in a quality-gate role — both get more careful under pressure, not faster.
- Don't let two agents of the same type talk to each other directly. They'll collapse into an echo chamber within a few turns.
- Cross-role decision rules go in user.md, not in the system prompt. Soul-level rules survive much better than prompt-level ones.
What Ganjiang does for a team isn't to fit one agent with a soul — it's to fit a group of agents with a set of cooperating souls. The leap from "one person" to "a team" is an orchestrator problem in engineering and a personality problem in Ganjiang.
“One agent's soul decides how it makes decisions. A team's souls decide whether they can collaborate.”