SWARMA setup
deploy your first learning agent swarm in 4 steps.
swarm
template
experiment
launch
1. name your swarm
what should your agent swarm be called and what's its mission?
lowercase, no spaces. this becomes your instance directory.
one sentence is fine. your agents will use this as their north star.
2. pick your setup
choose a swarm template and the harness that operates it.
content team
research, plan, write, analyze. full content pipeline with learning loop.
6 agents: researcher, growth lead, linkedin writer, twitter writer, visual designer, analytics
research team
deep research, analysis, and reporting across multiple sources.
3 agents: researcher, analyst, report writer
trading team
market scanning, signal generation, and trade analysis.
4 agents: scanner, signal analyst, risk manager, execution
blank canvas
start from scratch. add your own agents and define the flow.
0 agents: you decide
hermes recommended
full operator with shared memory, chat interface, and sub-agent delegation.
qmd memory / telegram+slack / soul.md personality / mcp tools
claude code
local terminal operator. agents work through your dev environment.
local files / mcp servers / terminal access
standalone
api-only mode. no operator layer, you orchestrate via http.
rest api / webhooks / byoo (bring your own operator)
custom endpoint
point at your own harness. any system that speaks http.
http post brief/result / adapter protocol
3. first experiment
give your swarm a hypothesis to test. this is how it learns.
one variable, one change. your swarm will test this and report back with a verdict.
agent
model
the order agents run in. "->" = sequential, "[a, b]" = parallel.
4. connect & launch
last step. connect your model provider and fire it up.