The connective layer becomes the product.

Anand Topu’s central reading of David Soria Parra’s keynote is not that models suddenly learned a new trick. It is that connectivity became the bottleneck.

A production agent must cross boundaries a demo can ignore: SaaS applications, internal data, shared drives, long-running workflows, identity providers, approval chains, and audit systems. Every connection creates both capability and risk. That is why MCP’s evolution—and the harness surrounding it—matters.

The scarce resource is no longer a tool call. It is a trustworthy path from intent to enterprise action.

A mature agent should not force every task through one connection shape. It should select the lane that preserves the most useful semantics with the least friction.

Loading every tool definition before the model understands the job is the connectivity equivalent of opening every drawer in a workshop. The agent sees more—and understands less.

Anthropic’s engineering guidance describes two complementary moves: discover tools only when needed, then use code to compose calls and filter intermediate results before they enter model context.

const notes = await drive.search(q)
const open = notes.filter(relevant)
await crm.update(open)
return summary(open)

A one-to-one REST wrapper can technically expose an API while still giving the agent a poor operating surface. Good tools compress low-level calls into meaningful work and preserve the interaction patterns a person would recognize.

If the tool surface would frustrate a human operator, it will probably confuse an agent too.

Since the Medium article appeared in April, official MCP updates have settled which parts of the 2026 connection layer landed in the core specification, which became extensions, and which remain separate efforts.

The official 2026 roadmap prioritizes transport scalability, agent communication, governance, and enterprise readiness. The release-candidate announcement showed how quickly several of those priorities moved into an implementable shape, and the July 28 release made several of them final.

MCP reduces the cost of connecting agents to systems. It does not decide who may act, what should be loaded, how risk is escalated, or what counts as complete. Those are harness decisions, and they expand with organizational complexity.

Everything above is architecture. Eve is what it looks like with a customer on the other end: Feniex’s assistant, in production, helping people choose equipment, make a buying decision, or get support for equipment they already own.

Eve runs on the Agentic Harness Deyaf packages: her knowledge, memory, doors, rules and learning loop. An Agentic Harness is everything around the AI model: what the assistant knows, what it remembers, where it meets customers, what it may do, and how your team teaches it. MCP answers one slice of that, how an agent reaches a system. The harness behind Eve answers the rest.

Field plate / 02Mood illustration

She is measured, not promised. On 100 fixed test questions, graded September 23, 2026:

As of September 24, 2026 · Feniex’s internal Eve 3.0 report · machine-judged and provisional. Eve’s report card, weak spots included ↗

When she cannot answer, a person answers once; the answer is checked, then taught. She is measured on a frozen test set before and after, and only explicitly approved lessons are published. The loop that makes her better is supervised, not training on every raw conversation.

Saved is not sent. Sent is not resolved. The harness lets her say only what the result proves.

Every connection in this briefing runs out of sight. A customer judges the whole system by the two doors in front of them, and for Eve those doors are small on purpose.