The hype around AI agents has reached a fever pitch. We’ve all seen the endless threads, demo videos, and funding rounds built on the promise of fully autonomous assistants that can handle entire workflows with minimal oversight. But in the real world? That promise continues to fall short.

Utkarsh Kanwat put it plainly in his recent piece, Why I’m Betting Against AI Agents in 2025 (Despite Building Them). While optimistic about AI’s potential, he’s clear-eyed about where things stand today.

“Meanwhile, the winners will be teams building constrained, domain-specific tools that use AI for the hard parts while maintaining human control or strict boundaries over critical decisions. Think less ‘autonomous everything’ and more ‘extremely capable assistants with clear boundaries.’”

We couldn’t agree more.

Fully Autonomous Agents Don’t Deliver What They Promise

The idea of an AI that can run your calendar, handle procurement, write code, or answer support tickets without human involvement is intoxicating. But autonomy has a reliability problem.

Agents hallucinate. They make assumptions. They struggle to reason across unfamiliar tools, shifting context, or exceptions—everything enterprise systems have in spades. And despite some impressive demos, the gap between “working in the lab” and “working in production” remains enormous.

Business leaders don’t want a clever prototype. They want consistent, auditable, high-confidence outcomes. That’s not what autonomous agents are good at.

Real Wins Are Coming From Orchestrated AI

At Contextual, we’re not chasing full autonomy. We’re building orchestration—tools that connect AI’s strengths to real data, structured processes, and defined decision points. It’s much more like directing an ensemble, than unleashing a soloist.

An orchestrated system isn’t just an agent running unchecked. It’s a network of AI capabilities working in coordination—each tuned for a role, operating within a known pattern, and accountable to observable outcomes.

This is what we mean by intelligent orchestration:

• AI that writes a first draft, but only within retrieved and validated source material
• Agents that take action, but only after human approval
• Flows that route exceptions to real people, not into black holes

Patterns, Not Promises

We’re often asked how to make AI more dependable. The answer isn’t just better models. It’s better scaffolding.

At Contextual, we’ve developed the idea of AI Patterns: repeatable frameworks for common business solutions like proposal generation, invoice processing, customer onboarding, or data classification. Each pattern defines what’s fixed, what’s flexible, and where AI should step in. This kind of design discipline is what keeps solutions from unraveling at scale.

Agents can be part of these patterns. But they are components—not the architecture.

Teams Don’t Need Magic. They Need Leverage.

Most organizations aren’t looking for agents that “think” like humans. They’re looking for leverage: faster approvals, fewer errors, less time lost to rote tasks.

That kind of impact doesn’t come from trying to make agents smarter. It comes from designing systems where agents play their part—alongside human judgment, systems of record, and defined flows.

When built this way, AI doesn’t just feel impressive. It becomes useful.

What Is It We’re Betting On, Then?

We’re not anti-agent. But we are anti-magic-thinking.

We believe the future belongs to teams who understand the difference between autonomy and orchestration—and who build systems where AI has guardrails, context, and accountability.

That’s why we built Contextual. At it’s heart, Contextual has always been a platform where AI agents don’t act alone, but as part of end-to-end orchestrated solutions. Businesses are using this platform right now, to design, deploy, and evolve real AI systems that solve real problems..

Autonomous agents might make the headlines. But orchestrated systems make the impact.

‍

Additional Reading:

Why I’m Betting Against AI Agents in 2025 (Despite Building Them) - Utkarsh Kanwat.
Kanwat makes a sharp, grounded case that autonomous agents won’t win in the short term—not because the vision is wrong, but because real-world execution is messy. His bet? The winners in 2025 will be teams building domain-specific, constrained tools that use AI for the hard parts while preserving human control and structure.

AI Can’t Do It All – Benedict Evans

Evans makes the case that while LLMs are impressive, they’re not turnkey magic. He argues for domain-specific application design, careful system integration, and—critically—orchestration of multiple components to make AI actually useful.

The Toolformer Paper: Language Models as Tools – Meta AI

This research explores how language models can select and orchestrate external tools—a key technical validation for orchestration approaches. The framing is important: the LLM isn’t the solution, it’s the interface to a network of tools.

The Trouble with Agents – Simon Willison

While a bit oder, this is a developer-level critique of agent-based AI. It’s highly readable and directly addresses failure modes in real-world usage. Good to include if your audience leans technical or skeptical.

‍