You can start by treating “agent swarms” as a structured workflow around Claude rather than a single monolithic setup, then layer in tools as needed.
Clarify research use‑cases
Given your work, the highest‑ROI swarm patterns are likely:
Literature deep‑dives and synthesis (multi‑agent “research → critique → synthesize”).
Theorem/proof exploration and counterexample search for DP/MDPs.
Code + experiment loops for simulations and numerical examples.
Writing pipeline: structure → draft → edit → reference/consistency check.
Pick one concrete use‑case (e.g. “multi‑agent pipeline to survey quantile MDPs and propose open questions”) as your pilot.
Use Claude’s built‑in agent teams first
Claude Code now has “agent teams” / swarm mode, where a lead agent plans and delegates to specialist agents in parallel.
To experiment with this without any infrastructure:
Enable the swarm/agent‑team view in Claude Code (UI exposes a team lead pane plus teammate panes where you can watch agents work).
Prompt the lead as a research manager, e.g. “Set up a team with: (1) literature‑scout, (2) formal‑methods checker, (3) code‑experimentalist, (4) writing‑editor; coordinate them to explore X and return a structured report plus code.”
Let Claude propose a plan and team, then refine roles and stopping conditions (“stop when you have at least 3 distinct approaches with pros/cons and verified references”).
This gives you hands‑on intuition about what multi‑agent coordination buys you before you invest in external frameworks.
Add orchestration when you hit limits
Once you want persistence, experiments, or integration with your own codebase, add an orchestration layer designed for Claude.
Good options:
Claude‑Flow: dedicated orchestration platform for Claude swarms with hierarchical “queen/worker” style control, consensus and shared memory.
Useful when you want longer‑running workflows, configurable swarm sizes, consensus rules, and integration with Claude Code via MCP.
oh‑my‑claude‑sisyphus: open‑source, Claude‑only multi‑agent framework built as a Claude Code plugin; defines agents and skills in simple config files.
Provides 10–12 named agents (Oracle, Librarian, Explore, etc.) and automatically handles sub‑task decomposition and “keep working until done” behavior.
For your workflow on Ubuntu/Python, these let you keep most logic local (git repo, experiments) while using Claude as the reasoning layer.
Design a minimal “research swarm” pattern
You can get a lot of benefit from a 3–5 agent pattern:
Lead/orchestrator: maintains the main question, decomposes tasks, checks alignment with your spec.
Literature agent: searches, extracts key definitions/results, builds a structured bib and summary.
Formal/derivation agent: works only on mathematical details, sketches proofs, checks steps.
Coding agent: implements examples and numerical tests in Python/JAX, under your repo constraints.
Reviewer agent: critiques assumptions, looks for gaps, suggests robustness checks.
In Claude Code swarm mode, you get versions of these roles via the built‑in team system; in Claude‑Flow or oh‑my‑claude‑sisyphus you’d define them explicitly in config.
You can start by treating “agent swarms” as a structured workflow around Claude rather than a single monolithic setup, then layer in tools as needed.
Clarify research use‑cases
Given your work, the highest‑ROI swarm patterns are likely:
Literature deep‑dives and synthesis (multi‑agent “research → critique → synthesize”).
Theorem/proof exploration and counterexample search for DP/MDPs.
Code + experiment loops for simulations and numerical examples.
Writing pipeline: structure → draft → edit → reference/consistency check.
Pick one concrete use‑case (e.g. “multi‑agent pipeline to survey quantile MDPs and propose open questions”) as your pilot.
Use Claude’s built‑in agent teams first
Claude Code now has “agent teams” / swarm mode, where a lead agent plans and delegates to specialist agents in parallel.
To experiment with this without any infrastructure:
Enable the swarm/agent‑team view in Claude Code (UI exposes a team lead pane plus teammate panes where you can watch agents work).
Prompt the lead as a research manager, e.g. “Set up a team with: (1) literature‑scout, (2) formal‑methods checker, (3) code‑experimentalist, (4) writing‑editor; coordinate them to explore X and return a structured report plus code.”
Let Claude propose a plan and team, then refine roles and stopping conditions (“stop when you have at least 3 distinct approaches with pros/cons and verified references”).
This gives you hands‑on intuition about what multi‑agent coordination buys you before you invest in external frameworks.
Add orchestration when you hit limits
Once you want persistence, experiments, or integration with your own codebase, add an orchestration layer designed for Claude.
Good options:
Claude‑Flow: dedicated orchestration platform for Claude swarms with hierarchical “queen/worker” style control, consensus and shared memory.
Useful when you want longer‑running workflows, configurable swarm sizes, consensus rules, and integration with Claude Code via MCP.
oh‑my‑claude‑sisyphus: open‑source, Claude‑only multi‑agent framework built as a Claude Code plugin; defines agents and skills in simple config files.
Provides 10–12 named agents (Oracle, Librarian, Explore, etc.) and automatically handles sub‑task decomposition and “keep working until done” behavior.
For your workflow on Ubuntu/Python, these let you keep most logic local (git repo, experiments) while using Claude as the reasoning layer.
Design a minimal “research swarm” pattern
You can get a lot of benefit from a 3–5 agent pattern:
Lead/orchestrator: maintains the main question, decomposes tasks, checks alignment with your spec.
Literature agent: searches, extracts key definitions/results, builds a structured bib and summary.
Formal/derivation agent: works only on mathematical details, sketches proofs, checks steps.
Coding agent: implements examples and numerical tests in Python/JAX, under your repo constraints.
Reviewer agent: critiques assumptions, looks for gaps, suggests robustness checks.
In Claude Code swarm mode, you get versions of these roles via the built‑in team system; in Claude‑Flow or oh‑my‑claude‑sisyphus you’d define them explicitly in config.