The freshest 24 hours were quiet on the architecture front — the only hard news was a consultancy launch — so today's issue also picks up what the past week shipped that we hadn't yet covered. The thread connecting all four items: the graph is leaving the code and becoming a document. Microsoft now defines multi-agent orchestration in YAML you can review, LangChain feeds its company analytics agent from context files versioned in GitHub, and the people who built the tooling are arguing about what to call the discipline. Four items today, dates marked on each.
Microsoft ships Declarative Workflows 1.0: orchestration as a reviewable document
Published July 23 and production-ready in both Python (agent-framework-declarative) and .NET: Agent Framework workflows defined entirely in YAML — agent invocation, conditional routing, loops, Power Fx expressions over workflow state, MCP and HTTP tool calls, human-in-the-loop approval gates, and checkpoint/resume for long-running executions. The stated payoff is organizational as much as technical: because “the workflow is a document rather than a call graph,” product owners and architects can read, diff, and version the orchestration without reading framework code. That is edge enumeration made auditable — the whole topology in one reviewable file.
LangChain rebuilds its own analytics on an agent-first data stack
Posted July 27: LangChain's three-person data team replaced their BI tool with a governed analytics agent and published the architecture. The interesting part is what surrounds the model — dbt semantic models define canonical metrics, “workspace guides” synced from GitHub document business logic and interpretation rules, endorsements mark which sources are trusted, and observability closes the loop on gaps. Results claimed: roughly 40× the request volume the team could handle directly, with about a third of the company using it inside 30 days. The lesson is pure graph engineering: the agent node was the easy part; the typed context and provenance edges around it did the work.
Harrison Chase and Sydney Runkle claim the graph-engineering name
In “3 Years of Graph Engineering with LangGraph” (July 22), the LangChain founders respond to the naming debate we covered in Issue #001 with a pointed opener: “Graph engineering isn't a new idea.” Their substantive claims deserve the attention more than the branding fight does: graphs encode domain knowledge by putting model reasoning in the right places with the right context; production agents need cycles for retries and revision, not clean DAGs; and — newly practical — entire agent runs can now sit inside a single graph node. They also concede the boundary: highly exploratory work like deep research fits an agent harness better than a drawn graph. That concession is Module 0's minimum-necessary-architecture rule, stated by the people selling the graph tool.
Cognizant launches an EMEA AI unit aimed squarely at the pilot-to-production gap
Announced July 28: a dedicated unit helping European, Middle Eastern, and African enterprises move agentic AI from pilots to production, structured in three tiers — Foundation (strategy, governance, prototypes), Accelerate (high-value use cases into production), and Transform (multi-agent delivery squads redesigning workflows end to end). EMEA president Manoj Mehta frames it as deliberately platform-agnostic: “We work across clouds, models and ecosystems.” When a systems integrator productizes the gap between agent demo and agent system, that gap is where the billable pain lives — and it is exactly the systems layer this discipline exists to teach.