In July 2025, Ramp shipped its first AI agents for controllers to enforce expense policies and flag fraud. In the same window, Stripe published its Agent Toolkit and Agentic Commerce Suite with Shared Payment Tokens scoped by time and amount. Coinbase’s engineering team described their enterprise agent architecture — a Tiger Team that shipped agents for Institutional support, Onramp onboarding, and Listing legal review with auditability and human oversight wired in. Three companies, three regulated verticals, no shared codebase. The architectures are structurally the same. That is not coincidence. It is convergent evolution — the signal that this shape of system is the only one that solves the problem.
The Pattern
Convergent evolution in biology: flight emerged independently in birds, bats, insects, and pterosaurs — not through copying, but through the constraint of aerodynamics. The same dynamic is visible in enterprise AI. Three companies, three different problems — payments, expense control, compliance — converged on the same four-layer architecture.
Layer one is the scoped role agent. Each agent has a narrow domain mandate and a hard authority boundary. Ramp’s controller agent approves only low-risk expenses below a threshold and escalates the rest with rationale, hitting 99% accuracy on its bounded slice. Stripe’s payment agents operate inside Shared Payment Tokens scoped to a single seller, time window, and maximum amount. Coinbase’s agents operate inside specific compliance lanes — never across them. Bounded scope is not a limitation. It is the design.
Layer two is the spawned task agent. When a Tier 1 agent hits something outside its mandate, it spawns a sub-agent for investigation rather than guessing. Reconciliation discrepancies, suspicious-receipt flags, transaction-screening anomalies — each becomes its own scoped task with its own evaluation. Layer three is the synthesis layer. Outputs from multiple agents flow into a coordinator that compares results and surfaces conflicts. Coinbase’s Case Grading Assistant compresses trainee review time from ~90 minutes to ~20 by running this layer cleanly. The synthesis layer is often where the most valuable work happens — not “the answer,” but the disagreement that requires a human. Layer four is memory and audit log. Every decision writes back to a structured store with full traceability. For Coinbase, this is regulatory non-negotiable. For Stripe, the substrate of fraud detection. For Ramp, policy learning. Memory is what turns one-shot decisions into compounding intelligence.
Why It Matters
The convergence is the diagnostic. When three regulated-vertical companies independently ship the same four-layer system — scoped role agents, spawned task agents, synthesis, memory — that shape is no longer a design choice. It is the production form of agentic AI in regulated verticals. Anything simpler does not survive enterprise procurement; anything more elaborate does not ship.
For founders, the signal is direct. If the architecture in your slide deck collapses any of the four layers — a single super-agent “handling everything end to end” with no synthesis or audit memory — you are building something the enterprise category has already evolved past. Buyers in regulated verticals already know what the working shape looks like. The startups that win in 2026 bring this same shape to verticals incumbents have not yet automated — clinical operations, claims adjudication, supply-chain reconciliation, AML tooling for non-crypto fintech.
The Charaka View
Manthan Intelligence’s analytical pipeline is the same architecture in a different vertical. Tier 1 persona agents own scoped analytical lenses — technology, returns, operations, ecosystem — and never reach outside their mandate. Tier 2 spawned agents handle deep dives when a lens flags something. Tier 3 synthesis surfaces tensions between lenses rather than averaging them. Tier 4 — the knowledge graph — logs every assessment so the next one is sharper. We did not copy Stripe or Ramp. We arrived here by watching simpler architectures fail in the same places everyone else’s failed.
The contrarian read for 2026 is that the architectural argument is over. The next wave of enterprise AI value is not in inventing new agent topologies — it is in applying the proven four-layer shape to verticals incumbents have not yet touched. The pattern has stabilised. The market opportunity has not.
This analysis draws on Ramp’s agent announcement via PRNewswire, Stripe’s Agentic Commerce Suite blog, and Coinbase’s engineering write-up on enterprise AI agents. Human editorial oversight applied.
This analysis is informational and does not constitute investment advice, a research report, or a recommendation to buy, sell, or hold any security.
Charaka Notes by Manthan Intelligence. Subscribe