TL;DR
- The biggest bottleneck in enterprise AI has shifted from generating code to orchestrating, validating, and controlling the cost of multiple AI agents working together.
- Research shows that optimizing the harness—the software that coordinates agents and tools—can cut token costs by 41% and execution time by 44%, regardless of which AI model you use.
- IBM's latest Bob platform update and Concho AI's new knowledge layer are two examples of infrastructure catching up to the orchestration problem.
- The enterprises that win won't be those with the biggest models—they'll be those who architect the smartest systems around the codebases they actually have.
The Quiet Revolution Nobody's Demoing at Their All-Hands
There's a particular kind of corporate AI announcement that gets everyone excited: a frontier model does something jaw-dropping, a demo video goes viral, a breathless press release lands in your inbox at 6 a.m. You know the drill.
But the most consequential shift happening in enterprise AI right now looks nothing like that. It doesn't come with a flashy benchmark or a dramatic chat transcript. It's happening in the plumbing—in the orchestration layers, the harness configurations, the semantic knowledge graphs, and the cost-analytics dashboards that most developers haven't even opened yet.
Welcome to the age of multi-agent orchestration. It's less glamorous than "vibe-coding." It's also considerably more important.
From Code Generation to Code Coordination
Not long ago, the core promise of AI in software development was simple: write faster. GitHub Copilot autocompletes. ChatGPT drafts functions. The bottleneck, as enterprises saw it, was generation—getting code onto the screen quickly.
That bottleneck has moved.
As IBM's July 2026 update to its Bob platform makes clear, the harder problem now is coordination: how do you get multiple agents to review, validate, and build on each other's output without spinning up runaway costs or producing inconsistent results? IBM Bob's new capabilities—parallel model-native tool calling, subagents for context management at scale, and a built-in analytics layer called Bobalytics—are a direct response to what enterprises are actually struggling with in production.
The results from early adopters tell the story better than any spec sheet. Jack Henry, a financial services firm, used the platform to accelerate legacy RPG development. Blue Pearl, meanwhile, completed what had been projected as a nine-month legacy modernization project in just three days. Three days. That's not a typo, and it's not magic—it's what happens when orchestration is done right.
"Enterprise development teams are shifting bottlenecks from code writing to code review and validation."
— IBM Bob platform update, July 2026
The 41% Nobody's Talking About
Here's where things get genuinely interesting—and a little humbling for anyone who's been obsessing over which frontier model to license.
Research published by Systima, summarized in a July 2026 CIO.com piece, introduces a concept called "The Harness Effect." The harness, in this context, is the orchestration software that coordinates models, tools, and workflows—the invisible scaffolding around your AI agents.
The finding? When researchers tested the same Claude Sonnet model on the same task in two different harnesses, they found dramatically different levels of baseline token overhead before a single user prompt was even sent. Same model. Same task. Different harness. Wildly different resource consumption.
By switching harnesses while keeping everything else constant, the researchers achieved:
- 38% reduction in token consumption
- 41% reduction in cost per task
- 44% reduction in execution time
- Comparable output quality throughout
Let that sink in. Nearly half your AI infrastructure costs and runtime could be a harness configuration problem—not a model problem, not a prompt problem. A plumbing problem.
And yet most enterprises have no visibility into this. CIOs aren't measuring what their harness sends per request. Cache behavior at scale remains a mystery. The meter is running, and nobody's watching it.
The Legacy Codebase Problem (Or: Archaeological Marvels Don't Scale)
Even if you nail your orchestration layer and optimize your harnesses, multi-agent systems face a third, deeply unglamorous challenge: they need to understand the code they're working with.
For greenfield projects and tidy modern repositories, this is manageable. AI agents can ingest the whole thing and get to work. But enterprise software is rarely a clean, well-documented monorepo. More often, it's the digital equivalent of a city that's been continuously rebuilt for forty years—layers of legacy logic, shadow functions written by developers who left the company a decade ago, and business rules embedded so deep that no single person fully understands them anymore.
Concho AI, which launched its flagship platform in July 2026, is tackling exactly this problem. Rather than asking developers to adopt new tooling, Concho inserts a semantic "fact layer" between the codebase and the AI assistant—a knowledge graph that understands not just what the code does, but why it exists and what business logic it encodes.
"What you're looking at is a combination of business-critical system and archaeological marvel," said Concho CTO Bruce Henderson in an interview with SiliconANGLE.
The platform uses what Henderson calls a "cognitive precompiler"—doing the heavy lifting of application discovery before any query is made, so that when a developer (or an AI agent) asks a question, it gets a fast, sourced, architecturally-grounded answer rather than a hallucinated guess.
The practical results are already visible. Clearwave, a private-equity-backed medical technology company, uses Concho to manage a codebase north of 12 million lines—a system that has, by all accounts, been through enough changes to qualify as a historical document. A small team of senior developers had been serving as the institutional memory of the entire system. Concho now handles routine knowledge transfer, freeing those experts from being consulted every hour and allowing the broader team to work with confidence.
Concho surfaces its insights through the Model Context Protocol, feeding tools like Claude with structured architectural understanding. The result: AI agents can work efficiently on systems they'd otherwise have no hope of navigating.
The New Competitive Landscape
Step back and the shape of the new enterprise AI landscape becomes clear. It's a three-layer problem:
- Orchestration — Can your agents coordinate effectively, in parallel, without spiraling costs?
- Harness optimization — Is your infrastructure generating unnecessary token overhead, and do you even know?
- Knowledge representation — Do your agents understand the codebase reality they're actually operating in?
Companies that are still focused primarily on which model to use are, at this point, optimizing the wrong variable. A mediocre harness running a state-of-the-art model will almost certainly underperform a well-architected harness running something a generation older—and cost more doing it.
This isn't an argument against keeping up with frontier models. It's an argument for treating orchestration architecture as a first-class engineering concern, not an afterthought.
What CIOs Should Actually Be Asking
If you're a technology leader evaluating your enterprise AI strategy right now, the questions worth asking have shifted. It's less "which model should we use?" and more:
- What does our harness send per request, on average?
- Do we have visibility into per-task token costs across different workflows?
- Can our AI agents access meaningful semantic context about our codebase, or are they flying blind?
- Are our agents set up for parallel execution, or are they waiting in line?
The enterprises pulling ahead aren't necessarily the ones with the biggest AI budgets or the hottest model subscriptions. They're the ones asking boring, operational questions about orchestration and then building rigorous systems around the answers.
The Revolution Looks Like Infrastructure
Enterprise technology has a long history of revolutions that only look obvious in retrospect. The database wasn't the exciting part of the early internet era—the applications were. But without solid data infrastructure, none of those applications would have survived contact with real users.
Multi-agent orchestration is the database moment for enterprise AI. The flashy layer is still the model, still the chat interface, still the generated output. But the durable competitive advantage is being built one layer down—in harness design, cost visibility, and knowledge architecture.
The companies that understand this are already compressing nine-month projects into three days.
The ones still chasing the next frontier model announcement might want to check their plumbing first.
Published in Stream · Dispatch #453 · July 15, 2026 · 7 min read.
Reply to paolo@mont3.ch - every email gets a human answer within 24h.