Two Axes, Not One Market

“AI agents for developers” isn't one market — it's two axes and four quadrants. Pick the wrong one and you'll swear the agent didn't work.

Most vendors are still selling “AI agents for developers” as if it's one market. It's two markets, split on two axes — and getting either axis wrong is what makes teams say “the agent didn't work for us.”

AXIS 1 · WHAT KIND OF WORK?
PAVED ROAD
Known workflow, run at scale.
The SDLC exists — review conventions, deploy pipelines, a bug backlog. The job is to industrialize it: more of it, faster, with governance, audit, and a review queue that doesn't drown the team. Optimization is the game.
FRONTIER
New workflow, tight loop.
The workflow doesn't exist yet — you're building software, or a category of software, that hasn't been built before. Primitives aren't settled; abstractions are up for grabs. The human needs direct control and tools that don't opinionate them into a shape that hasn't proven itself.

Sell a Paved Road tool to a Frontier team and you strangle the exploration that made them valuable. Sell a Frontier tool to a Paved Road org and you get three senior devs going 10x while the rest of the team can't plug in — the productivity gain evaporates at the team boundary.

AXIS 2 · WHERE DOES THE AGENT RUN?
LOCAL
On your machine.
Terminal, IDE, direct filesystem access, a tight feedback loop, human-in-the-loop by default. Great for exploring — harder to govern at team scale.
REMOTE
On someone else's.
Cloud sandbox, ephemeral VM, hosted runtime. Parallel by default, no laptop dependency. Governance, RBAC, and audit are easier because the execution surface is centralized — at the cost of a slower loop.

These axes are related but not the same. Remote execution tends to correlate with Paved Road (because governance and parallelism scale better in the cloud), and Local tends to correlate with Frontier (because tight loops beat round-trips when you're exploring). But you can genuinely have all four combinations, and they demand different products.

How the frontier is actually splitting

LOCALREMOTE
PAVED
ROAD
Local pipelines
Sparse — remote usually wins here
Headless CLI in CI
Scheduled agents on-box
Claude Code --headless in pipelines
Cloud orchestration
Where the enterprise money is going
Augment Cosmos
Cognition Devin
OpenAI Codex Cloud / Web
Google Jules
GitHub Copilot Workspace
FRONTIER
Builder territory
Where new abstractions get made
Claude Code (interactive)
Codex CLI
Grok Build
Cursor / Aider
Remote frontier
Niche — but growing
Cloud dev sandboxes
Ephemeral prototyping envs
GPU-heavy exploration
Pair-across-timezones setups
  • Augment Code made the loudest bet: Paved Road + Remote. They sunset the VS Code and JetBrains extensions and pushed everyone to Cosmos — a cloud runtime with an Expert Registry, shared memory across sessions, and prebuilt experts for code review, PRs, testing, and incident response. They kept the Context Engine reachable as MCP and shipped Auggie CLI, but the primary product is unambiguously the enterprise factory. Individual devs on PyCharm were told, politely, that they're not the customer.
  • Anthropic anchored Frontier + Local. Claude Code is terminal-first, built around builder primitives (CLAUDE.md, skills, subagents, hooks, MCP), and leans “supervised autonomy” on purpose. The Paved Road pieces they ship — Dreaming, Outcomes, dynamic workflows, the Agent SDK — are sold as primitives for builders to assemble, not a finished production line. Cowork extends this into non-dev workflows without giving up the builder-primitive posture.
  • OpenAI runs both plays in parallel. Codex CLI (open-source, Apache-2.0) is the Local Frontier surface. Codex App / Cloud / Web with approval gates, RBAC, and OS-level sandboxing is the Remote Paved Road surface — and it's what earned them Leader status in Gartner's enterprise coding-agent Magic Quadrant.
  • xAI started Local Frontier with Grok Build — local-first, no source sent to their servers — and is bolting on Paved Road pieces as they come (Skills, Connectors, /goal long-running mode).
  • Cognition's Devin is pure Remote + Paved Road: give it a ticket, get a PR. Cursor is Local Frontier being pulled toward Paved Road. GitHub Copilot started Local Frontier, moved into Remote Paved Road (Workspace), and now covers both awkwardly.

Why this matters for your evaluation

  • If you're a buyer, figure out which quadrant you're in before you evaluate tools. And notice that most enterprises need surfaces in more than one quadrant — the mistake isn't buying multiple products, it's buying one and pretending it covers all four.
  • If you're building these products, pick your quadrant. Trying to serve all four with one surface is how roadmaps become incoherent and pricing pages read like a group project.

The model gap is closing every quarter. The winners will be the ones who know which quadrant they're serving on which surface — and who resist the temptation to blur it.

Which quadrant is your team actually working in — and are your tools sitting in the same one?

Share this post
Tags
Archive
Because of AI, We Are Brute-Forcing Everything
Cheap compute removed the pressure that used to force the smarter way — and it's starting to cost AI itself the most.