Jonathan Ingles: Long week — but honestly the reading I did for this made it longer. You see the gap I kept circling back to?
Cole Bryant: The 23 versus 5 thing? Yeah. Bro, I — I actually wrote it down on paper because I didn't trust myself to keep it straight in my head.
Jonathan Ingles: Twenty-three percent of enterprises running multi-agent AI at scale right now in 2026. Five percent generating actual value. That's not a maturity curve. That's a structural problem the industry has collectively decided not to name.
Cole Bryant: And multi-agent is — I mean, we should say what we actually mean — these aren't single-model queries, this is agentic AI, systems doing multi-step autonomous work, planning, coordinating fleets of specialized agents on complex tasks. And even with all that, 95% can't show the outcomes.
Jonathan Ingles: The one public response that acknowledges this directly is Orange. French telecom. €600 million value target, explicit outcome-first philosophy. The industry response to the same data is to ship more agents faster.
Cole Bryant: Wait — €600 million is the target, not the result yet, right? That's an important distinction, dude.
Jonathan Ingles: It's the stated target. Which is still more honest than the companies scaling with no target at all.
Cole Bryant: Okay, that's — yeah, fair. So what we're actually trying to figure out today is whether the industry is confusing shipping agents with deploying something that works. And why the one company saying 'measure first' is the weird one in the room.
Jonathan Ingles: The part that breaks the 'measure first' framing open — and this is the thing nobody's naming — is that the industry hasn't just skipped validation. It's structurally replaced it. The new argument is: you don't validate models upfront, you route between them at runtime. Every single request gets assigned to whatever model is cheapest and appropriate in that specific moment. That's not a bug. That's the product.
Cole Bryant: Wait — like, each task gets a different model? Not just one model doing everything?
Jonathan Ingles: Think of a hospital staffing dispatcher. Instead of hiring one reliable surgeon and verifying credentials before they ever touch a patient — you just call whoever's available and cheapest at the moment the patient arrives. No prior vetting. That is runtime routing. That's the analogy. Model quality becomes a live cost-optimization decision, not a design decision.
Cole Bryant: Okay that's — I mean, that's a little terrifying but I also see why an engineering team looks at that and goes 'yes, this is efficient.' Because ACRouter, the thing that actually implements this, beat fixed premium-model strategies by 2.6x on cost. That number is — bro. 2.6x. That's not a rounding error.
Jonathan Ingles: Released in 2026. Open-source. The mechanism is a Context-Action-Feedback loop — the router itself acts as an agent, it builds memory of which models worked on which tasks over time. No pre-deployment quality gate required. And frankly, that's why Microsoft Azure AI Foundry, AWS Bedrock AgentCore, and Google Vertex AI are now competing on fleet orchestration rather than single-model quality. The economic incentive is right there in the 2.6x number.
Cole Bryant: So InfoQ literally — wait, no, they called it a software architecture discipline. Not an AI problem anymore. The reliability gets engineered at the orchestration layer, not per model.
Jonathan Ingles: Which sounds sophisticated. It is sophisticated. But what it's also doing is giving every engineering team a rational justification for never running a pre-deployment validation pass. Why would you, when the router handles it live and it's 2.6x cheaper?
Cole Bryant: The incentive is genuinely compelling though — I can't pretend it isn't. If I'm running an engineering team and someone shows me that number, I'm shipping.
Jonathan Ingles: That's exactly why it spreads. The 2.6x is real. The problem is what you've quietly agreed to when you accept it — quality is no longer a gate. It's a variable.
Cole Bryant: But — okay, quality as a variable, that's the framing the Orange story is supposed to fix, right? Like, that's the take circulating. 'Orange proves you can do this responsibly at scale.' Dinootoo, 100,000 employees, 80% of their global workforce. People are citing that like it's the counterproof.
Jonathan Ingles: 80% adoption of an internal GenAI chat environment. That's a deployment metric.
Cole Bryant: Wait — say that distinction again because I think people are going to miss it.
Jonathan Ingles: Dinootoo launched in 2023. The stated purpose was reducing shadow AI — employees using unsanctioned tools outside approved systems. By 2026, 80% of Orange's workforce is inside the approved environment. That means Orange won the governance problem. It does not mean Orange generated competitive, customer-facing value. Those are different things.
Cole Bryant: No, okay, that's — I mean, I want to push back a little because containing shadow AI is actually real? Like, if your employees are running sensitive customer data through random consumer tools, that's a live liability. Orange built a secure environment, got 80% of its workforce inside it. That's not nothing.
Jonathan Ingles: It's a risk-reduction story. I'll give you that. But the €600 million figure — that's a target. Stated. Not independently verified. So the proof point being held up as 'outcome-first works' is actually: we contained a security risk, and we have a number we're aiming for.
Cole Bryant: Which is — yeah, that's genuinely narrower than how it's being sold. Human-in-the-loop design, iterative rollout, all of that is real discipline. But discipline on the governance side isn't the same as proof on the value side.
Jonathan Ingles: The proof point is genuine. It's just smaller than the claim. And the healthcare data — that one actually cuts both ways in a way that breaks this whole framing open. We should get to that.
Cole Bryant: Wait, how does it cut both ways? Like — both directions at once?
Jonathan Ingles: OncoAgents. Multi-agent AI platform, neuro-symbolic architecture — meaning it combines neural language models with structured knowledge graphs for auditability. Prospectively evaluated across 3,804 consecutive oncology patients over twelve months for clinical trial matching. That is one of the largest real-world validations of an agentic AI system in any high-stakes domain. It works. The evidence is there.
Cole Bryant: Wait — 3,804 patients, prospective, twelve months? That's not a pilot.
Jonathan Ingles: No. And the same domain — oncology, same year, 2026 — is where healthcare researchers published a seven-principle evidence-informed framework arguing you need systematic evidence review before clinical deployment. So the field that has the best proof it works is also the field demanding the most rigorous pre-deployment validation. Both things are true simultaneously.
Cole Bryant: That's — okay, that's actually kind of wild? Because you'd expect — I mean, if the evidence is there, why are the researchers in that same domain the ones saying slow down?
Jonathan Ingles: Because the stakes are legally and ethically catastrophic if it fails. Validation-first is only being practiced where failure is not survivable. Everywhere else — enterprise workflows, internal tooling, productivity agents — the industry ships and routes. The discipline exists. It's just rationed.
Cole Bryant: So the proposed fix — and this is where context engineering comes in, right? Deloitte, KPMG, both cited it in 2026 enterprise research — the idea that you design the entire informational environment the agent operates in. Not just the prompt. The relevance of the data, the sufficiency, the provenance. Frame it like the agent's operating system.
Jonathan Ingles: Consulting firms citing it is not enterprises implementing it. And frankly, context engineering requires exactly what continuous-deployment culture discarded — clear objectives before you ship, pre-deployment testing, measurable outcome definitions baked in from the start. That's the governance framework. Most teams don't have that infrastructure.
Cole Bryant: No but — okay, imagine a clinical informatics team at a mid-size cancer center. They see the OncoAgents data, 3,804 patients, they want it. And the seven-principle framework is sitting right there. Does the governance actually keep pace with how fast agent fleets are being deployed right now?
Jonathan Ingles: That's the structural question. And the honest answer is: in oncology, maybe. Because the liability forces the discipline. In every other domain, the fleet ships before the framework catches up. That gap is what to watch.
Cole Bryant: That gap is what I keep — I mean, I can't figure out whether the 95% failure rate eventually forces a reckoning or whether Microsoft Azure AI Foundry, AWS Bedrock AgentCore, Google Vertex AI, all of them, just... price it in. Because those platforms make money on agent volume. More agents, more compute, more routing calls. The failure rate is almost — it's almost neutral to them financially?
Jonathan Ingles: That's the whole thing. Orange's outcome-first model is a direct threat to that business. Fewer agents, fewer API calls, less revenue for the orchestration layer. The incentive to let it spread simply doesn't exist for the platforms that dominate deployment. It's not a conspiracy. It's just math.
Cole Bryant: So the question I'm left with — and I genuinely don't know the answer — is whether any customer failure rate is high enough to change that math. Like, does 95% ever become the number that breaks the model? Or does it just become the expected cost of shipping fast?
Jonathan Ingles: I don't know. And I think that's the honest place to leave it.
Cole Bryant: Good talk, bro. I mean that — genuinely useful to think through.
