Orchestration

Air Traffic Control
for your agent fleet.

This isn't a co-pilot sitting next to you. You're the air traffic controller — coordinating multiple autonomous agents without babysitting each one.

Multi-Agent Fleet

Specialized agents for specialized tasks

Run a security reviewer, a style checker, a performance analyst, and a documentation writer — all from one console. Each agent has its own model, system prompt, tools, and knowledge base.

Goal-Oriented

Define objectives, not step-by-step instructions. Agents figure out how.

Full Observability

Real-time status dashboard. See what every agent is doing right now.

Complete Audit Trail

Every decision, every tool call, every response — logged and searchable.

Agent Dimensions

Fine-tune agent behavior with per-dimension parameter overrides.

[Screenshot: agent fleet dashboard showing multiple agents with active/standby status indicators]
Chain PipelineInputRaw dataStage 1ClassifierIntent detectionStage 2ExtractorEntity parsingStage 3GeneratorOutput craftedOutputAggregation: Structured result with chain provenance! prompt template → auto-filled per stageError → re-route to fallbackEach stage is a reusable building blockChain stages can be agents, tools, workflows, or sub-chainsOutput of stage N → input of stage N+1 with auto-prompt generation
Chains

Pipeline your prompts like Unix pipes

Chains run agents in sequence. Each stage gets the output of the previous one. Use aggregation strategies — majority vote, best-of, concatenate — to synthesize multi-agent responses into a single output.

Aggregation Strategies

Random, majority-vote, best-of, or concatenate — you choose how responses combine.

Loopable

Chains can loop with configurable termination conditions.

Step Tracking

See progress at every chain level. Know what stage each agent is at.

Declarative Specs

Define chains as versioned TOML files. Infrastructure-as-code for AI pipelines.

Dialogues

Let your agents talk to each other

Dialogues are multi-agent conversations. Run them in parallel (all respond to the same prompt) or as a chain (each sees the prior response). Define cycles, termination conditions, and bridge prompts.

Parallel Mode

All agents respond simultaneously to the same input. Collect the best answer.

Chain Mode

Agents respond in sequence. Each builds on the last.

Configurable Cycles

Define how many rounds of dialogue to run. Let the conversation evolve.

Auto-Termination

Consensus detection, keyword triggers, timeouts, token budgets — set your own kill switches.

Dialogue ConfigurationsParallel ModeAll agents respond simultaneouslyUserAExpertBCriticCSummarizerAll respond in one turn ~3x fasterChain ModeSequential, each sees prior outputInASees inputBSees A's outputCSees A+B outputsOutCoherent narrative, context buildsChoose your dialogue strategyParallel: Fast iteration, independent responses. Chain: Coherent narrative, context builds.Cycle count: repeat dialogue N times until consensus. Hybrid: parallel groups within chain stages.Configure max cycles, stop conditions, and aggregation strategies per dialogue group.
Hook System ArchitectureTrigger Pointson_messageon_responseon_branchon_errorBinding LayerHook → Agent routingFilter conditionsPriority orderingTimeout configObserving AgentsMeta Agent 1Log monitorMeta Agent 2Quality checkMeta Agent 3Re-route fallbackDispatchedActions→ Intercept message→ Modify context→ Terminate branch→ Re-route agent→ Trigger hook→ Log auditHooks enable self-observing AI systemsAny trigger point can bind to any meta agent. Meta agents observe, intercept, or modify the primary flow.Use cases: moderation, quality gating, cost control, audit logging, automatic re-routing on failure.
Meta Agents

AI agents that manage AI agents

Meta agents observe, annotate, and act on conversations. Attach a meta agent to any hook point — message send, tool call, prompt input — and give it scoped permission to intervene.

Observation Scopes

Single message, full conversation, or entire workspace. You set the boundaries.

Action Grants

Warn, block, reroute, or transform. Meta agents can intervene at any hook point.

Topic Tracking

Visual divergence detection. See when conversations drift from their intended path.

Intercept Rules

Define policies as cards. Test with a harness before deploying.

Ready to command your fleet?

From single-agent chats to multi-agent dialogues — Bonsai scales with you.