John Jayasankar
Résumé (opens in a new tab) Email me

JJ-SYS-06Agent platform3 min read

A reusable control plane for five enterprise agent systems

One-off agent integrations do not scale. A shared control plane for context, typed actions, approval, and evals made the next agent a reuse problem, not a prototype from zero.

  • Quantile · LSEG
  • Production
  • 2025
  • Lead Product Manager
Next: Pre-Trade Margin SimulatorAll work

Agent Platform & Controls

One plug path · five systems · scope expands when evals say so

5systems, one control model

Schematic · same interfaces · one shared control spine

Phases1-5 · select to hold

Reading the schematic interfaceFive enterprise systems on one control plane, passing context through MCP servers and domain APIs, typed actions, deterministic services and a human approval gate, and finishing at shared evaluation baselines.

Agent PlatformBroken01/07

What was broken

Individual AI workflows were becoming one-off integrations. Production scale needed a repeatable way to expose domain context and actions while keeping behavior bounded, observable, and reviewable.

Product decision

Standardize human-supervised architecture: reusable MCP servers, domain APIs, Pydantic-typed actions, deterministic services, HITL gates, and shared evaluation / QA baselines.

In plain terms

  • Five agent systems on one pattern beat five bespoke prototypes
  • MCP and domain APIs expose context; typed actions limit what an agent can attempt
  • High-consequence steps stay in deterministic services with human approval gates
  • Shared evaluation and QA baselines make “production-ready” a repeatable path, not a one-off review

What I owned

  • Standardized the product and control pattern across five enterprise agent systems
  • Defined the human-supervised architecture: interfaces, typed actions, HITL gates, and eval baselines
  • Treated agent development as a reusable product capability rather than a sequence of prototypes

Control model

  • Typed actions constrain what an agent can attempt
  • Deterministic services keep high-consequence logic outside the model
  • Human approval gates consequential workflow steps
  • Evaluation and QA baselines create a repeatable path from prototype to production

Role & tags

Lead Product Manager. Standardized the product and control pattern across five enterprise systems.

  • Reusable interfaces
  • Typed actions
  • Deterministic services
  • Evaluation baseline

Impact

New agent workflows reuse proven interfaces instead of starting from zero. Agent development became a product capability, not a sequence of prototypes.

Next systemPre-Trade Margin Simulator0 to 1 · originated and shippedRead the case

Also seeCompression Setup AgentIncident Triage AgentCross-Currency Compression

1-7 beats · j / k · ← → adjacent · y link · b brief