What Happened
Production agent architectures are converging on orchestration patterns that separate routing from reasoning, while new tools address cost control, evaluation gaps, and stateless deployment. Meta’s thin orchestrator delegates domain logic to specialists instead of stuffing context into a single prompt. Spending caps are emerging as authorization primitives for autonomous agents with cloud access. SIT-Eval exposes how aggregate metrics hide context-dependent failures in conversational agents. Meanwhile, developers are building agent-facing infrastructure from unexpected primitives: Org-mode as task coordination, Lambda-based repo scoring, and fragment-hash state storage for agent-generated documents.
Why It Matters
Orchestration is the new bottleneck. Meta’s production failure wasn’t solved by bigger models or better prompts—it required architectural separation of concerns. The thin orchestrator pattern inverts the common approach of centralized reasoning, showing that reliability comes from limiting what the router knows, not expanding it.
Cost control is a security boundary. As agents gain payment access, spending caps become authorization primitives that must be enforced outside the agent’s control plane. Hard limits on cloud provisioning and API spending prevent runaway loops from becoming financial incidents.
Evaluation frameworks lag production reality. SIT-Eval demonstrates that models with identical aggregate scores diverge sharply on context-dependent tasks. Multi-criteria, turn-by-turn evaluation reveals orchestration challenges that single-number benchmarks obscure.
Key Trends
Orchestration patterns favor delegation over centralization. Meta’s thin orchestrator enforces reliability checks and routes work to specialists rather than embedding domain knowledge in the router. This reduces hallucination by limiting the orchestrator’s decision surface.
Plaintext formats are becoming agent coordination layers. Celorga exposes Org-mode’s task management as MCP tools, decoupling the data format from the editor. Agents can read, write, and trigger actions from structured plaintext without breaking 15 years of compatibility.
Stateless architectures eliminate persistence overhead. Fragment-hash state storage lets agents generate shareable documents without touching a backend. Compression and client-side rendering turn URLs into self-contained artifacts.
Supply chain security requires workflow gatekeepers. RepoPilot scores repositories before cloning, addressing the 7,600 fake repos shipped in 2026. Serverless scoring pipelines run in under two minutes without standing infrastructure, showing how to insert security checks into developer workflows.
Authorization primitives are shifting to resource limits. Spending caps as hard boundaries, scoped tokens, and pre-approval workflows define what agents can do without human intervention. The control plane must be external to the agent.