AI Agent Protocols in 2026: Choose the Boundary Before the Standard
MCP, A2A, and ACP did not converge into one universal protocol. Their current roles make sense once agent systems are divided by interaction boundary.
Artificial Intelligence
Models, agents, learning, reasoning, evaluation, and the technical or conceptual consequences of machine intelligence.
10 published materials across notes, projects, and publications.
Primary in this field
Work whose central question belongs to Artificial Intelligence.
Agentic Resource Discovery Still Needs a Verification Boundary
ARD can help agents find capabilities across organizations, but discovery metadata must remain outside the trusted execution boundary.
Choose Agent Infrastructure by the First Constraint That Can Break It
Agent stacks should be selected layer by layer, with explicit contracts for state, tools, memory, evaluation, and recovery.
Chain-of-Thought Prompting Can Improve Answers; It Is Not Proof
Chain-of-thought prompting can improve some multi-step answers, but a plausible rationale is not evidence that an answer is correct or faithful.
Coding Agents Work Best When Verification Is Cheaper Than Reimplementation
Terence Tao's restored mathematical applets show where coding agents have real leverage—tasks whose outputs remain cheap for an expert to verify.
Cross-field connections
Work centred elsewhere that materially uses or informs Artificial Intelligence.
Token Usage Is an Input Cost, Not an Engineering Outcome
Counting tokens repeats an old measurement mistake unless AI-assisted work is evaluated through delivery, quality, and user outcomes.
Paper Diagrams Are Maps, Not Proof Certificates
Dependency diagrams can make mathematical papers easier to navigate, provided their inferred structure is not mistaken for verified proof evidence.
Agentic End-to-End Testing Still Needs an Independent Oracle
Slack's agent-driven E2E experiments expose a useful new execution model, but adaptation must remain separate from the authority to declare success.
Selecting Fewer MCP Tools Is Not the Same as Enforcing Least Privilege
GitHub MCP tool filtering reduces context and accidental capability exposure, but real least privilege also requires identity, authorization, and effect controls.
AI-Obsidian Service
Local indexing and retrieval-augmented generation for an Obsidian vault, with REST APIs and SSE progress streaming.