Superset

Cursor, GitHub Copilot, Windsurf, and the rest were all designed around a single human paired with a single agent — one chat window, one editor, one diff at a time. That shape was necessary at the time: the underlying coding models still needed close supervision, manual edits, and frequent human overwrites for anything non-trivial. Which is exactly why these products were all built as forks of VS Code with the text editor as the primary surface and the agent as an assistant in the sidebar.

Then the models got good. Claude Sonnet 4.6 / Opus 4.7 (1M context), GPT-5 / Codex, and Antigravity can now write whole features, review their own PRs, run their own tests, and ship without a hand-holder. The 5–10× productivity gain everyone talks about only shows up when you run several of them at once — and the new top 5–10% of builders, the tokenmaxxers, have already shifted: instead of waiting idly while one agent works, they spin up many in parallel and review the outputs as they finish.

But running several at once is where the friction lives. You're juggling terminal tabs, learning tmux, setting up git worktrees by hand, switching branches, re-orienting yourself every time you come back to the agent you started twenty minutes ago, then reviewing diffs across scattered sessions. The VS-Code-fork architecture wasn't designed for this — it was designed for one human typing.

Most engineers respond by either (a) running one agent at a time and idling while it works, or (b) building bespoke internal infrastructure. The Fortune 500 dev orgs took option (b): Ramp built Inspect,^[7] Stripe built Minions,^[8] Spotify built Honk.^[9] OpenAI coined the term harness engineering to describe the practice — and used it internally to ship a beta product of roughly 1M lines of code with zero manually-written source code.^[4] Each of them is now publishing engineering posts about how essential this layer is — and how expensive it was to build.

Model reliability crossed the bar. The inflection point was December 2025: Claude Opus 4.7 (1M context) shipped, GPT-5/Codex matured, and Anthropic and OpenAI's coding harnesses leveled up around them — Skills, subagents, slash commands, hooks, persistent memory. The experimental pattern of one human running many agents in parallel tipped into a production-grade workflow within a quarter.

Model-provider pricing punished the API-wrappers. In a six-week stretch this spring, Anthropic cut third-party harnesses off from Claude subscriptions, then reinstated access via a capped Agent SDK credit pool billed at full API rates — a step-change cost increase for any tool that wraps the API.^[17] Superset's CLI-agnostic architecture sidesteps the squeeze: agents run as subprocesses against the user's existing Claude Code, Codex, or Antigravity subscription. The same structural insulation would absorb the next pricing shift, whichever provider triggers it.

The flagship orgs went public. Ramp,^[7] Stripe,^[8] and Spotify^[9] have all disclosed in-house builds of the software-factory model in the last six months — making the architecture legible to enterprise buyers and putting Superset in position to be the off-the-shelf fleet console.

The MVP that came out of the YC hackathon: a desktop IDE built around cmd+N spawning isolated git worktrees, with any CLI coding agent (Claude Code, Codex CLI, OpenCode, Antigravity CLI) running inside each one. Notifications when each agent finishes. GitHub-style diff review across sessions. One-click open in VS Code, Cursor, JetBrains, or Xcode. Linear integration to fire agents at backlog tickets directly.

v1 found product-market fit fast — 7k+ DAUs growing 50%+ w/w, 10.8k GitHub stars, front-page HN in five months.^[14] But worktrees don't scale: shared CPU, RAM, and disk space across many parallel agents, overlapping builds, messy envs and secrets, all bound to one machine. The worktree model worked beautifully — until it ran into the physical limits of the laptop it was running on.

The pivot: move the workloads to the cloud. v2 shipped in early May 2026, layering the missing infrastructure underneath the IDE — and notably, the team kept the v1 IDE shipping, the Discord swarming, the DAUs compounding 50%+ w/w, and the enterprise pilots closing the entire time they were building it.

Remote Workspaces. Ship a session to the cloud; continue working from anywhere; share with teammates. Solves the laptop-bound limitation that internal background agents at Ramp and Stripe both name as their core unlock.^[7][8] Start on your laptop, continue on your phone, scale to as many agents as you need.

Automations. Schedule agents to pick up tasks for you later — or fire them on events: a new issue opened in Linear or GitHub, a Sentry error detected, a failed CI run, an inbound Slack message.

Superset CLI. Gives agents new superpowers and unlocks programmatic workflows — turning Superset itself into a primitive other agents can call.

MCP refresh. Works with v2 + an expanded toolset. Integrates across the apps where product and engineering work happens — Slack, Linear, GitHub, and more. Message @Superset in Slack to kick off a new task without ever opening the IDE.

Mobile · coming soon. Talk to your codebase from anywhere. Kick off a task on your phone walking to a meeting, review the diff on the train, ship from the couch. Same Remote Workspace, different surface.

The loop. Take a task. Spawn a sandboxed agent with the right tools and codebase context. Write code. Run CI. Open a PR. Review and iterate. Repeat in parallel — dozens, hundreds, thousands of times. The human's job has narrowed to architecting the loop and reviewing what comes out the other end, not typing the diff. Ramp's Inspect, Stripe's Minions, and Spotify's Honk are three implementations of the same loop.^[7][8][9]

The pillars. Three primitives make the loop work: isolated compute (each agent's blast radius is sandboxed), event routing (work flows in and out automatically), and governance (humans review, approve, and shape what merges).^[3] Superset's v2 maps cleanly to all three. Together they're what moves humans from on the line to on the loop.

The economic argument. Every team that has done this in-house has spent 6+ months and senior engineers to build it. Most companies will not. Superset's bet: nobody actually wants to build their own factory machinery — they just want the IDE that ships with it, so their engineers can step off the line on day one.

Near term — tokenmaxxers. Individual engineers paying $200+/month for Claude Code or Codex who would happily run more agents if the friction were lower. 7k+ DAUs growing 50%+ week-over-week, 10.8k stars, and front-page HN inside five months — the natural-demand signal.^[6][14]

Mid term — agentic coding teams. Small-to-mid engineering orgs (5–200 engineers) standardizing on a parallel-agent workflow as a team. Multi-seat enterprise deals are already in motion with Ramp and Thrive Holdings. The revenue stack is per-seat IDE plus the org-level platform features — Remote Workspaces, Automations, shared memory, admin controls, SSO/RBAC, audit logs — that get pulled in once an IC's workflow becomes a team workflow.

Long term — Software Factories. Dozens to hundreds of agents triggered upstream from Linear, GitHub, Slack, and on-call — running in sandboxed cloud workspaces and shipping PRs continuously. Ramp Inspect, Stripe Minions, and Spotify Honk are early in-house implementations.^[7][8][9] Every team that wants this without spending six months building it is a future Superset customer.