Section 01 · the 2026 consensus
Supervisor won. Swarm did not.
multi-agent topologies · what works in production
Swarm is press-release-grade. Supervisor + a few specialists is what teams actually ship.
Section 02 · the stack split
Provider SDK + portable orchestrator.
leaf agents use the best per-model SDK · orchestration is portable
The "framework war" is overblown. Pin LangGraph or Mastra. Leaves change quarterly.
Section 03 · harness over framework
Anthropic's pattern: batteries included.
Deep Agents / Anthropic two-agent harness (Mar 2026)
A harness ships the loop. A framework ships the primitives. In 2026, ship a harness.
Section 04 · tool · skill · subagent
Three primitives. Pick by eval.
Will Steuk · Anthropic · three ways to decompose a prompt that outgrew itself
Three primitives. The golden set decides which cut wins - promote / revert by eval delta, not intuition.
Section 05 · mixture-of-agents
Route across models and tools.
Kay Zhu · Genspark · not Mixture-of-Experts (model-internal) - Mixture-of-Agents is model+tool routing
MoE picks experts inside the model. MoA picks model + tool combos outside it - a routing surface, not a pipeline.
Section 06 · OSS framework picks 2026
Pick by stack, not by hype.
all OSS, all production-ready · pick by language & team
LangGraph + Deep Agents for Python. Mastra for TypeScript. Pydantic AI for typed back-office. All OSS.
Section 06b · the blackboard
Three primitives. All multi-agent needs.
shack-blackboard · KV + pub/sub + locks · classic blackboard, modern MCP
A KV. A topic log. A lock table. All N-agent coordination patterns reduce to these three.
Section 07 · vollko OSS
The framework-agnostic primitives.
· · ·
Build the AI-native firm