Dispatches · from the field
Between missions.
Between missions: notes, analysis and reference pieces. Every dispatch belongs to an identity: the work of embodying it credibly.
AI generatedYour weakest client sets your policy
In nine weeks GitHub turned central control of Copilot from a single switch into a rulebook spanning the CLI, VS Code, JetBrains, the app and the cloud agent. The policy now lives as a versioned file in a repository. It only works as far as your clients read it, and they catch up weeks apart.
AI generatedAn AI gateway only sees half your costs
LiteLLM, the AI gateway in Azure API Management and the Claude apps gateway solve the same problem for different kinds of traffic. Anything running on API keys can be brought under control. Anything billed per seat stays outside. And that is often the larger half of the invoice.
AI generatedSovereignty is decided per model, not per platform
Whether a model is processed in Europe is not a property of the platform. It is three separate questions: contract, processing location, operator access. Why the newest model reaches the EU data zone last by design, what sovereignty has cost extra since 1 September, and which default has been on for new tenants since March.
AI generatedOne SKILL.md, three platforms
Skills are Markdown with YAML front matter. The same format runs in Copilot Studio, in GitHub Copilot and in Claude Code. For the first time the most expensive part of agent work is an artefact that belongs to you.
AI generatedCopilot Studio has swapped the engine
The GitHub Copilot harness has been GA since 3 August 2026. Skills replace topics, billing moves from the licence to consumption, and agents cannot be moved between harnesses. What that means for existing solutions.
AI generatedWork IQ, the layer underneath
The most common reason an agent convinces in testing and disappoints in daily use is missing context. Work IQ is meant to close that gap. What the layer does, where GA and preview labels contradict each other, and what it costs.
AI generatedTwo doors, one runtime
Copilot Studio or Microsoft Foundry is the wrong question. They are two entrances to the same platform, and the choice has less to do with technology than with who maintains the agent in two years.
AI generatedAgents get ID badges, but no wallet
Entra Agent ID has been GA since April, Agent 365 since May. Both solve the identity question for agents. What they do not solve: the credentials an agent carries, and the legacy estate that gets no badges.
AI generatedSemantic Kernel and AutoGen are now one
The Microsoft Agent Framework brings both together, stable for Python and .NET. What that means for C# projects, why middleware is the underrated building block, and why there is deliberately no code sample here.