DevGPT
Cloud-Native AI Coding Agent Platform
An internal platform that enables J.P. Morgan Chase engineers, product & designers to delegate software and prototype development tasks to AI agents — from feature implementation to deployment, running securely within the firm's infrastructure at enterprise scale.
Stateless by Design: Context & Cache at Scale →Transport only — connection lifecycle, auth, routing, Redis Streams proxying. All harness logic lives exclusively in the worker image. No exceptions.
All agents are stateless. State lives externally in Redis or Temporal's durable execution engine. Temporal's replay guarantees are a first-class design constraint.
Temporal manages intention state continuity across agents and sub-agents — surviving crashes, evictions, failures — while the router/dispatch layer remains stateless.
Our harness is the single source of truth for agent loop logic. IDE plugins are thin context capture and UI bridge layers only. Porting creates divergence risk.
A dedicated memory service backs the harness with EFS-scoped memory files, Redis, and PostgreSQL/pgvector — retrieval fuses vector similarity with BM25 via Reciprocal Rank Fusion, and an LLM scope classifier tags insights team- vs. project-scoped before Temporal-orchestrated consolidation promotes them to long-term memory.
Pre-warmed pod pools eliminate cold-start latency at dispatch. Remote sandbox controller pods sit ready behind Istio Ambient Mode — sidecar-free mTLS and L4 policy enforced at the node level, keeping the data plane out of the agent container entirely.
All automated patch branches are reviewed by a supervisor agent before surfacing to human owners. Teams retain merge authority to release/master. Automation proposes; humans decide.
DevGPT powers TLM — a firm-wide JPMC initiative for automated software maintenance. Agentic workflows are currently delivering Java/Moneta Boot migrations on a dedicated maintenance branch, with a roadmap to a harness-driven agent loop and broader language coverage across Python, Golang, TypeScript, and Terraform. A supervisor agent reviews proposed changes and routes them to the owning team for merge to release/master. TLM operates across the full firm while the platform team remains AWM-chartered.
Agents execute untrusted code in sandboxed subprocesses via a purpose-built remote sandbox controller, which leverages bubblewrap for lightweight, kernel-level process isolation — no privileged containers, no separate runtime service. Kubernetes NetworkPolicy egress controls (deny-all except DNS) sit behind corporate HTTP proxy containment, and an Identity Proxy brokers AWS credentials to sandboxes via an init-container authenticated by Kubernetes TokenReview — the harness never holds credentials directly. Sub-2-second dispatch target with warm pod pool strategies via Karpenter NodePool and KEDA ScaledObject.
A high-throughput event platform (OpenTelemetry ingest, MSK/Kafka streaming) processes 50K+ events/minute from IDE plugins and LLM proxies, with per-response token attribution including cache read/write tokens. The platform is evolving to a CQRS + CDC architecture on Aurora Serverless v2 — separating real-time AI attribution writes from analytics and reporting reads.
In 2025, led a 2-week migration of DevGPT from AWM's shared AWS org into a dedicated AWS org via a Route53 weighted DNS cutover, to support firm-wide scale — completed with a 30-minute planned maintenance window, zero data loss, no user-facing failures.