Swarm Intelligence

Bootspring's Swarm Intelligence layer enables coordinated multi-agent orchestration where agents don't just run — they collaborate through shared state, vote on decisions, and dynamically reorganize based on workload.

Topologies#

Choose the right agent organization for your task:

Agent Roles#

• Queen — Coordinator that manages topology and task dispatch
• Worker — General-purpose task executor
• Specialist — Domain expert (security, performance, architecture)
• Reviewer — Code review and quality validation
• Monitor — Health monitoring and metrics collection

CLI Usage#

Swarm Templates#

Pre-built configurations for common workflows:

• code-review — Star topology with security, performance, and style reviewers
• deploy — Hierarchical with build, test, and deploy specialists
• security-audit — Mesh topology for parallel vulnerability scanning
• feature-build — Hierarchical with parallel implementation and testing
• research — Mesh topology for parallel exploration

Deploy a template:

Vector Memory#

Agents store and recall knowledge using HNSW-indexed vector similarity search:

• Decisions — Record and recall past architectural choices
• Patterns — Learn from successful code patterns
• Errors — Remember and avoid past mistakes
• Trajectories — Replay successful goal-to-deployment sequences

Search is sub-5ms for 50,000+ entries using approximate nearest neighbor.

Goal Planner (GOAP)#

The Goal-Oriented Action Planning system decomposes high-level goals into executable task DAGs:

1. Define a goal and current state
2. The A* planner finds the optimal action sequence
3. Independent actions are grouped for parallel execution
4. Feed the plan directly into a swarm for execution

Built-in actions cover the full dev lifecycle: analyze, design, implement, test, review, security scan, deploy.

Agent Specialization Learning#

Agents improve over time. The engine tracks per-agent, per-task-type performance:

• Success rate — How often each agent succeeds at each task type
• Speed — Average duration per task type per agent (EWMA-smoothed)
• Smart routing — Task dispatch prioritizes agents proven best at each type
• Persistence — Specialization data survives swarm save/restore

Swarm Persistence#

Save and restore entire swarm states — agents, tasks, specializations, and execution logs:

Cost-Aware Routing#

Configure per-model costs and the engine automatically factors cost into agent selection:

Swarm Replay#

Every swarm records an execution log for debugging and analysis:

API Endpoints#

Dashboard#

The swarm dashboard at /swarm provides 7 tabs:

• Overview — Live topology graph (Canvas), metric cards, SSE event feed, spawn/submit forms
• Agents — Agent table with role, status, completion stats, remove button
• Tasks — Task queue with priority badges, status tracking, agent assignment
• Consensus — Run votes with configurable quorum, see results with bar charts
• Memory — Semantic search across agent memories, store new entries, view stats
• Metrics — Task/agent breakdown charts, performance bars, uptime/throughput
• Plan — GOAP planner with visual DAG of parallel execution groups

Multi-Model Deep Integration#

Assign specific AI models to specific agents, benchmark their performance, and configure automatic failover:

Model-per-Agent Assignment#

Model Benchmarking#

Compare models on actual swarm tasks to find the optimal assignment:

Fallback Chains#

Configure ordered lists of models with automatic failover when a provider goes down:

Per-Agent Cost Tracking#

Local Model Support (Ollama)#

Run swarms offline using local models via Ollama:

Supported models: Claude (Opus/Sonnet/Haiku), GPT-4o, GPT-4o Mini, Gemini 2.0 Flash, Ollama local.

Models Dashboard#

The multi-model dashboard at /swarm/models provides 6 tabs:

• Overview — Model health status, aggregate stats, provider breakdown
• Assignments — Assign models to agents, view current assignment table
• Benchmarks — Model rankings by task type, recent benchmark results
• Chains — Create and manage fallback chains with visual chain display
• Costs — Cost summary by model and agent, top spender analysis
• Ollama — Configure local model, health check, enable/disable

Automatic Memory Recall#

Passive per-turn memory recall — every agent message is searched against the vector memory graph and relevant memories are automatically injected into context. No explicit tool calls needed.

Memory types: decisions, patterns, errors, solutions, preferences, context. Extraction happens automatically from agent responses using pattern detection. Near-duplicates are filtered. Importance decays over time with access-based boosting.

Swarm File-Conflict Detection#

Tracks which files each agent has read or written. When an agent writes to a file another agent has read, a conflict event is emitted for real-time notification.

Structure-Aware Search#

Enhanced code search that returns the enclosing function, class, or method for each match. Saves context tokens by letting agents understand file structure without reading entire files.

Returns: { filePath, line, matchText, enclosingScope: { name, kind, line, signature }, language }. Supports TypeScript, JavaScript, Python, Rust, and Go.

MCP Tools#

Three MCP tools are available for AI assistants:

• bootspring_swarm — Swarm lifecycle, agents, tasks, consensus, topology
• bootspring_memory — Vector memory search, store, trajectories
• bootspring_plan — GOAP goal decomposition and planning