Anthropic Launched Managed Agents. We'd Been Running That Architecture for Four Months.

Four weeks ago, Anthropic launched “Managed Agents” — multiagent orchestration with rubric-driven self-improvement and outcomes loops. 331K views in 13 hours.

We’ve been running this architecture for four months.

Since February 2026, Manthan Intelligence has operated a multi-dimensional Analytical Council where specialised agent lenses — technology, macro, competitive, regulatory — run structured analysis batches against a knowledge graph of 100,000+ entities. Each lens evaluates independently. Results get cross-validated. The system grades its own output against human decisions and recalibrates.

Anthropic productised the pattern. Here’s what we learned building it from scratch — before the productised version existed.

Lesson 1: The orchestration layer is not the hard part.

Getting agents to call other agents is engineering. Getting them to disagree productively is design. Our biggest breakthrough wasn’t the orchestration — it was structuring how different analytical lenses surface contradictions rather than averaging them out. When your technology lens says “strong moat” and your macro lens says “regulatory headwind incoming,” the system needs to escalate that tension, not smooth it.

Lesson 2: Rubric-driven self-improvement only works if the rubrics evolve.

We’ve graded 800+ scorecards against real investment outcomes. Current weighted accuracy: 65.3%. The rubrics that worked in March don’t work in May — not because the models changed, but because the market context shifted. Static rubrics produce static blind spots. We rebuild calibration baselines monthly.

Lesson 3: The human’s job changes — and that’s the uncomfortable part.

When the system gets good enough, the GP’s role shifts from “analyse deals” to “interrogate the system’s analysis.” You stop being the analyst and start being the quality controller, the taste-maker, the person who asks “what did the system miss?” That transition is psychologically harder than any technical challenge we’ve faced.

Anthropic’s Managed Agents will make this architecture accessible to thousands of teams. That’s good — it validates everything we’ve been building. The moat isn’t the architecture anymore. It’s the 100,000+ entities in the knowledge graph, the 800+ calibrated scorecards, and four months of compounding institutional memory.

The infrastructure is becoming a commodity. The intelligence built on top of it is not.

---

More on the builder journey — Read more at getmanthan.com

Mayank Mathur | Founder, Manthan Intelligence | GP, Tavaga Fund