The connective tissue of the agent economy is about to get re-plumbed. The Model Context Protocol — the interoperability standard that lets AI models securely reach into calendars, databases, and internal tools — receives a significant update next week, one that changes how it handles session IDs so servers can operate more easily at a larger scale, according to TechCrunch. The changes might not be noticeable to end users, TechCrunch notes, but they could make a big difference in how the ecosystem develops. Founders, and the people writing their checks, should be paying close attention.

The timing could hardly be better for one startup in particular. Arcade, a two-year-old company that helps AI agents securely connect to and act on tools like Gmail, Slack, and Salesforce, raised $60 million in June on a specific thesis: most AI agents don't fail because the underlying models are weak — they fail because the infrastructure around them isn't ready yet. This protocol update is that thesis getting validated at the spec level.

The Session ID Problem, Explained by Someone Living It

The official spec for the new MCP version has been public since May, but Arcade founder Nate Barbettini offered an unusually clear explanation of the changes on Monday. At the heart of the update: session IDs, the small tokens servers use to remember that a request belongs to an ongoing conversation.

Under the current design, a client like Claude introduces itself, receives a session ID, and presents it on every subsequent request. That works until you try to run it the way real companies run things:

"

"Picture a real deployment. You're running a server for millions of users, behind a load balancer whose entire job is to route each request to whatever server in the farm is free, sometimes in a different region. Now every one of those machines has to know about a session ID that some other machine handed out. It's not impossible, but it's a serious pain, and it fights the load balancer instead of working with it."

In plain terms: today's MCP assumes one server remembers you, while enterprise deployments spread traffic across dozens of machines that don't share memory by default. That mismatch has been a genuine headache for anyone running MCP servers at scale — and a quiet brake on larger production rollouts.

The Harness Wars Are Still Wide Open

The protocol fix is one front in a much bigger fight over what software should surround the models. Ars Technica surfaced a genuine architectural split this week between two credible camps.

On one side: Anthropic's Cat Wu, head of product for Claude Code, who argues models are improving so quickly that building opinionated scaffolding around them is a losing game. Claude Code deliberately maintains a "lean harness" — the software layer that decides what a model sees and what actions it can take — and avoids pre-indexing codebases by default. "Going by the evals, we don't see a measurable change," Wu said of structured-context approaches. "And I think we generally lean more toward shipping a leaner harness with fewer opinionated tools and just letting developers add their own if they want."

On the other: Augment Code, whose VP of Engineering Vinay Perneti defends the semantic route — pre-indexing repositories with embeddings, a retrieval model, and a vector database to surface conceptually relevant code, rather than the grep-based approach used by Claude Code and OpenAI's Codex. The field now includes Google's Antigravity, open-source OpenCode, Cursor, and more.

Here's what excites me as someone who watches where companies get built: when well-resourced incumbents publicly disagree about fundamental architecture — and neither side has conclusive evals — the layer is still up for grabs. That is startup-shaped territory.

Agents Watching Agents Is Now a Product Category

LangChain published two pieces Monday that, read together, sketch the next layer of the stack. The first introduces IssueBench, an internal benchmark built to evaluate LangSmith Engine — an agent whose entire job is finding and fixing issues in other agents by analyzing their traces, per the LangChain Blog.

IssueBench comprises 15 tasks spanning three domains and 15 issue categories, using synthetically generated traces with known, labeled failures, sandboxed on Harbor and scored against hidden ground truth. The insight buried in the design: flagging a failing trace isn't enough. Engine has to cluster failures by root cause without creating noise, merge duplicates without losing detail, and assign categories correctly — because a hallucination, a silent tool error, and a feature gap route to different fixes and different owners. Agent supervision now has its own measurement science.

The second piece argues that governance has become runtime infrastructure. LangChain frames the LLM gateway as "the runtime control plane for enterprise AI," enforcing policy on every model call, tool call, and agent hop, with a five-part operating model: Govern, Decide, Protect, Observe, Assure (LangChain Blog). The drivers are familiar to any founder selling into the enterprise: unpredictable token spend, uptime requirements prototypes never had, and regulation — the EU AI Act carries significant penalties for noncompliance. Notably, LangChain also observes that open-source models are narrowing the quality gap "particularly when paired with a tuned agent harness," at a fraction of frontier prices. Governance, cost control, and the harness debate are the same conversation.

The Bill Comes Due at the Storage Layer

One more hidden cost surfaced this week. A sponsored analysis in the Harvard Business Review — worth flagging as sponsored — lays out the physics: a dozen typed words balloon into a 40,000-token workload once policies, session history, and retrieved documents attach. At the largest companies, knowledge bases run to 100 petabytes, and the key-value cache that makes responses reusable scales with concurrent users, not data volume. DRAM is too expensive to hold it, hard drives too slow to serve it, and GPUs frequently sit idle waiting on data movement — dragging out time to first token while the meter runs.

And there's a human design variable, too: a separate Harvard Business Review piece warns of growing evidence that AI use can erode critical thinking — a reminder that how we build these systems determines whether they sharpen or dull the people inside them.

What This Means for the Money

The value chain is re-sorting in real time. The model layer still gets the keynotes, but this week's news cluster — a protocol scaling fix, an unsettled harness war, benchmarks for agent supervisors, governance gateways, storage economics — points the same direction: the durable companies of the agent era are being built in connectivity, context, evaluation, and control. Arcade's $60 million is the visible tip of that repricing.

I've covered enough cycles to know the pattern: first the gold rush, then the fortunes made selling picks, shovels, and maps. The founders building unglamorous infrastructure nobody thanks them for — until everything breaks — are the ones I'd watch.

What to Watch

Three things. First, next week's MCP rollout and how quickly the major client makers adopt the new session handling. Second, published evals in the lean-versus-rich harness fight — whoever produces credible numbers first will shape the narrative, and the term sheets. Third, governance spending, as the EU AI Act's significant penalties for noncompliance turn consistent policy enforcement into what LangChain calls a material business risk.

The first act of the agent era was about making models smart. The second act — the one starting now — is about making agents work: reliably, affordably, at scale. That fight for survival, against constraints nobody chose, is where real companies get built.