Hook
Big Tech is pitching AI agents as the next major digital interface. Read the marketing copy and you get three words repeated like a mantra: trust, interoperability, efficiency. No company names. No revenue numbers. No latency figures. No end-to-end task success rates.
That absence is the signal.
When a sector has real traction, the press release leads with a number. When it has a narrative, it leads with adjectives. Every bull market has a word it hides behind. In 2021 it was "composable." In 2025 it's "agent." I spent 2021 watching DeFi protocols pitch "trustless" and "composable" while liquidity mining subsidies propped up 90% of their TVL. The adjectives were identical. The exit was identical too. When the emissions stopped, the "users" vanished โ because they were never users. They were yield farmers renting TVL by the block.
This is the same pitch in a new wrapper. And if you're a trader, the wrapper is not where the money is. The protocol underneath is.
Context
Let me be precise about what's actually being sold.
The claim: the interface is shifting. Today, a human opens an application and operates it. Tomorrow, a human states intent, and an agent operates the software on their behalf. That's the entire thesis. It's a real shift, and it matters โ but not for the reason the pitch decks say.
The shift matters because the interface is the distribution channel. Whoever owns the default entry point owns the traffic. This is not a technology argument. It's a land-grab argument. The browser did this. The App Store did this. Now the agent runtime is trying to do it, and the prize is the same: the right to tax every transaction that flows through you.
Notice what Big Tech stopped saying. They stopped pitching the model. Model capability is now table stakes โ the frontier models all clear the bar for single-step tasks. The competition moved up the stack, to the orchestration layer and the communication protocol between agents and their tools. That's where the lock-in lives, because that's where switching costs accumulate.
For crypto readers, this should sound familiar. This is exactly the OP Stack versus ZK Stack debate, and I've written this argument before. The technical merits were always secondary. The real question was: who convinces more projects to deploy on their chain first? Standards are won by adoption, not by elegance. The best protocol does not win. The most-adopted protocol wins.
One more thing the copy avoids: jobs. The industry uses the neutral word "impact" because the honest word is displacement. Customer service is already the first casualty โ high-volume, structured, and measurable, which makes it the easiest to automate and the easiest to justify cutting. That's not a forecast. That's a budget line already moving. And the "new jobs" counterargument โ agent trainers, prompt engineers, agent ops โ is real but small. You cannot reabsorb a displaced workforce of thousands with a role that scales in dozens.
Core
Here's the order flow. Two protocols are fighting to become the default.
Anthropic shipped MCP โ Model Context Protocol โ in late 2024, open source. It became the de facto standard for tool calling. OpenAI and Google adopted it. Then Google shipped A2A โ Agent-to-Agent โ in April 2025, with 50-plus launch partners, betting that agent-to-agent coordination, not agent-to-tool, is the layer that matters. Microsoft's Copilot stack is MCP-compatible and rides Windows and Office distribution โ the strongest install base in software.
This is a standards war, and standards wars are zero-sum until someone forces interoperability. MCP and A2A compatibility is still unresolved. That's not a footnote. That's the whole trade. Whoever's protocol gets embedded in the default runtime collects rent on every downstream call. Whoever doesn't gets routed around.
Let me put numbers on the actual bottleneck, because nobody in the marketing copy will.
Agent reliability compounds badly. A single-step task at 90% success sounds great. Chain ten steps together and you're at 0.9^10 โ roughly 35%. This is the reliability cliff. It's the reason most enterprise agent pilots never leave the POC stage. I've audited enough systems to know the demo is always clean and the production path is always dirty. Multi-step agents accumulate error, hallucinate tool calls, lose state between sessions, and blow past latency budgets. A conversational response takes a second. A ten-step agent workflow with tool calls can take thirty seconds or more. That is not an interface. That is a batch job wearing a chat window. Treat the demo as marketing and the success rate as the truth.
The enterprise math is where this gets exposed. Customer service is the fastest displacement target โ high volume, structured, measurable. Office productivity is the strongest enhancement target. Software engineering sits in the middle: heavy augmentation, partial replacement. But the adoption curve is gated by one number nobody publishes: average contract value and renewal rate. If agents deliver a 35% task success rate in production, the renewal never comes. That's the "POC graveyard" โ pilots that demo well and never scale. You don't see it in the press release. You see it in the renewal data two quarters later.
Now the compute side, because this is where the tradeable asymmetry sits.
Agent workloads invert the inference math. A simple chat turn consumes a few hundred tokens. An agent task โ multi-turn reasoning, tool calls, result integration โ consumes 10 to 100 times that. If agents scale, inference demand goes non-linear. Not training demand. Inference demand. That's a different supply chain: inference-optimized silicon, low-latency cloud, edge inference. If you were positioned for training capex, you're positioned for the wrong curve.
And the unit economics are brutal. Every agent task burns tokens the user never sees. If cost per task doesn't fall fast, the business model doesn't clear. I run an autonomous trading agent. I allocated $100,000 into compute and audits before it made a dollar. I know exactly what inference cost does to a P&L. It eats the margin before it eats the competition.
Map the positioning while you can still see it. Microsoft has the strongest distribution โ Windows and Office are the default surface for knowledge work, and Copilot rides them. Google has search, Android, and Chrome, plus the A2A protocol. Anthropic has MCP and the developer mindshare. OpenAI has the consumer habit and the strongest brand. Meta plays the open-source angle to commoditize everyone else's moat. The scoring is not on model benchmarks. It's on install base and protocol adoption.
Contrarian
Everyone is watching the models. The money is in the middleware and the trust layer.
Two blind spots.
First: the orchestration layer. The frameworks that route between MCP, A2A, and proprietary toolchains. That's the defensible position, because you don't have to pick a winning protocol โ you bet on the layer that has to speak all of them. In the L2 wars, the winners weren't only the chains. They were the bridges and the RPC providers that served every chain. Same structure, different asset class. When you can't call the winner, own the chokepoint that every winner has to pass through.
Second: the trust layer. Big Tech puts "trust" first in the pitch, and that ordering is a confession. When an AI only generates text, a mistake is a wrong sentence. When an agent executes operations, a mistake is a wrong transaction. The risk escalates from saying the wrong thing to doing the wrong thing โ and the blast radius scales with the permissions you grant.
The specific vector is prompt injection. An agent that reads your email or browses a web page can be fed malicious instructions by that content. There is no fundamental fix. The industry consensus is that it's unsolved. Add overreach โ an agent with account access that mis-fires โ and data exfiltration across every application the agent touches. The trust triad is identity, authorization, and auditability: who is acting, what they're allowed to do, and what's on the record afterward. Note that "auditability" is a polite word for "someone can reconstruct what happened after it went wrong." That's not security. That's forensics.
Here's the crypto connection nobody is pricing correctly. Verifiability is the one thing crypto actually adds to the agent stack. On-chain identity, scoped permissions, immutable audit trails โ these are not features you bolt onto a centralized runtime after the fact. They're the native primitive. If agents are going to hold keys and move money, the trust layer has to be verifiable. That's a real market, and it's the same lesson as my Parlay Protocol trade: security flaws are market inefficiencies, and they're mispriced until the exploit lands.
But I need to be blunt. 90% of what's branded "AI agent plus crypto" right now is a centralized product with a token stapled on. I watched this exact thing with "Bitcoin Layer 2s" in 2024 โ Ethereum projects rebranding for a narrative the actual Bitcoin community never acknowledged. Same playbook here. Rebrand, raise, exit. The real infrastructure is a handful of teams. The rest is packaging. Distinguish them by one question: does the token do anything the centralized version couldn't? If the answer is no, you're buying a narrative, not a protocol.
Takeaway
Watch two signals, not the pitches.
One: MCP and A2A convergence. If they merge into a compatible standard, the middleware layer compresses and the winners are whoever shipped adoption first. If they fragment, the multi-protocol orchestrators win, and you get protocol blocs โ exactly like the L2 landscape. Track the compatibility announcements, not the demo videos.
Two: the first major agent security incident. It's coming. An agent with real permissions will be prompt-injected into a real loss. When it happens, the trust layer reprices overnight โ and it reprices toward verifiable, on-chain enforcement, because that's the only version that survives contact with an actual adversary.
Big Tech is selling you an interface. The interface is the marketing. The protocol, the compute curve, and the trust layer are the trade. Position on those, and the adjectives stop mattering.