Insights
RSSWriting on multi-agent systems, the work AI can be trusted with, and what we're learning from building and running them in production.
Latest MCP, A2A, AP2, KYA, ATP, AGNTCY. The protocols showing up in every AI conversation, explained without hype or vendor framing.
Anthropic's financial-services repo is interesting because of the artefacts, deployment paths, data connectors, and human checkpoints around the agents.
Most founders are running into two problems at once. Their context is scattered across fifteen tools, and what they can ship in a week is capped by how many hours they can sit at a keyboard. There is a quieter answer that is starting to look like a real edge.
Many growth-stage teams have done an AI pilot that worked and somehow never reached production. The reasons are rarely technical. The five patterns we see, and the three pilot-design questions that change the outcome.
Most teams treat EU AI Act compliance as a documentation exercise. For agent systems operating in regulated environments, that is the wrong frame. The requirements that matter are built in or bolted on, and the difference shows.
Agent systems that work at launch quietly degrade by month six. The named decay modes, the operating discipline that prevents them, and why custodianship is what Tier 04 Manage is actually selling.
A concrete implementation reference for the routing, state, contracts, observability, and recovery patterns that make multi-agent orchestrators hold in production.
In a multi-agent system, the agents are not the product. The orchestration layer is. What it does, how to architect it, and why the framework choice matters less than taking the layer seriously.
Most teams that commission agent systems should have done six months of operator enablement first. Skipping that step is the dominant reason six-figure AI projects underperform. An argument for the unsexy first step.
Most teams pick the wrong first automation. Here is the ranking framework we use in every Diagnosis: leverage, frequency, tractability, reversibility, and the trap of starting with the loudest pain.
Building multi-agent systems badly is worse than not building them at all. The six disciplines that decide whether yours survives, and the two failure modes you only see in production.
Single agents work. Multi-agent systems work differently, and that difference is the whole game for operational functions.
No posts match your search.