defract › blog

Paradev vs Claude Code: parallel agent runners after Conductor

2026-11-10 7 min read

On November 4, 2026, Conductor announced it was joining Meta. Cloud shutdown date: November 20. New sign-ups stopped the same day. Within the week, Paradev went MIT (open source, desktop free forever) and Claude Code shipped its dynamic workflows feature to general availability across all paid plans. For developers who built their parallel-agent workflow around Conductor, the question is now which of these two replaces it, and whether that is even the right frame.

Paradev and Claude Code's native parallelism are solving adjacent but different problems. Paradev is a workspace that manages multiple CLI coding agents side by side, regardless of which model or agent they run. Claude Code's parallel capabilities (worktrees, subagents, dynamic workflows) are built into the tool you are already using, no extra app needed. The choice between them is less about feature lists and more about where you want the complexity to live.

Paradev: MIT workspace for parallel CLI agents

Paradev started as Superset, rebranded in October 2026, and went fully open source under the MIT license on November 8, 2026. The desktop app is free; running agents locally will never require payment. That position is a direct play for Conductor's user base, which ran on a free local app with an optional paid cloud layer.

The core proposition is breadth: Paradev supports 20+ CLI coding agents in a single workspace, including Claude Code, Codex, OpenCode, Cursor Agent, Gemini CLI, GitHub Copilot CLI, Grok, and Kiro. Each task gets its own git worktree, branch, and terminal, so agents work on separate tasks without stepping on each other. The workspace UI shows all running sessions, diffs, and agent statuses from one panel. For teams that run multiple agent tools, or want to compare outputs across models, the routing layer is a genuine capability most single-agent tools do not offer.

Cloud workspaces (agents running when your machine is off) are available on paid plans: Team at $60 per user per month (150 compute-hours, 5 cloud environments) and Ultra at $200 per user per month (500 compute-hours, 10 environments). The desktop is BYO-subscription on every plan, so Paradev adds no token markup on top of your existing Claude, Codex, or other model spend.

Where Paradev stands out:

  • Multi-vendor agent routing. Run Claude Code and Codex on the same codebase from one workspace. Teams evaluating models side by side, or standardizing on a mix, have one surface to manage it from.
  • MIT / free desktop. No subscription required to run parallel agents locally. The open-source move lowers adoption risk and invites community contributions.
  • Scale ceiling. 100+ simultaneous agents for large sweeps (test generation, documentation passes, security audits across a codebase). The visual workspace makes it manageable at that volume.
  • Cloud execution on paid tiers. Agents run server-side when your laptop is closed. That is a capability the native Claude Code parallelism options do not cover for all usage patterns.

Where Paradev adds friction:

  • macOS only (Linux experimental). Windows is not supported. Teams with Windows developers cannot run Paradev uniformly across the team.
  • No process layer. Paradev dispatches agents efficiently but does not define what they build or in what order. There is no scope checkpoint, no approval gate between stages, no design pass. It manages the work you already know you want; structuring what that work should be is still on you.
  • Cloud execution costs add up. Local Paradev is free, but cloud workspaces at $60-200 per user per month is a meaningful addition on top of your model API spend, depending on team size.

Claude Code: parallel built into the tool you already have

Claude Code ships three parallel execution models, at different capability levels. Manual worktrees and subagents are generally available. Dynamic workflows went GA across all paid plans in mid-2026. Agent Teams remains experimental (opt-in via environment variable).

Worktrees are the manual approach: you create separate git worktrees and run independent Claude Code sessions in each. No extra app needed. Each session has full context, its own branch, and runs independently. The coordination layer is you: briefing each agent, tracking what each is working on, managing merge order. This is what most parallel-agent practitioners already do, and Claude Code 2.x made worktrees first-class (clear worktree commands, improved isolation). The ceiling is a few agents before coordination becomes the bottleneck.

Subagents (GA) let Claude Code spawn child agents mid-task to handle parallel subtasks and return results to the parent context. The orchestration is code-level: the parent agent coordinates by writing instructions for each child. Subagents share your Claude subscription and bill against the same token budget. This is the right model for tasks where one agent can intelligently delegate parts of a problem rather than you manually briefing each one.

Dynamic workflows (GA, May 2026) are the most powerful option. When a task needs large-scale parallel analysis, Claude writes a JavaScript orchestration script that fans the work out across up to 1,000 parallel subagents, runs adversarial verification on findings, and returns only the final verified output to your session. The orchestration is entirely automated: Claude plans it, writes it, and runs it. You describe what you want; dynamic workflows figure out the fan-out strategy. The use cases it was built for are codebase-wide bug hunts, security audits, and large optimization sweeps, tasks where scale matters and automation is the point.

Agent Teams (experimental, off by default) adds a different capability: multiple independent Claude Code instances working as a coordinated team, with an agent panel for directing and messaging teammates. Enable with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. Still behind a flag nine months after launch.

Where Claude Code's native parallelism stands out:

  • No additional app or subscription. Dynamic workflows, subagents, and manual worktrees all run in the tool you already pay for, against your existing Claude subscription.
  • Dynamic workflows automate the orchestration itself. Claude writes the fan-out strategy, not you. At large scale (100+ subagents), that is a meaningful reduction in coordination overhead.
  • First-party capabilities land here first. New Claude Code features, model access, and primitives arrive in the native tool before third-party wrappers can adopt them.

Where Claude Code's parallel options add friction:

  • Claude-only. No multi-vendor routing. If you want to compare Codex and Claude on the same task, or run a mix of agents, native Claude Code parallelism does not help.
  • Dynamic workflows are fully automated. Claude writes the orchestration script. For use cases where you want to control exactly which agents run which tasks, manual worktrees are the option, and those require manual coordination at scale.
  • No visual workspace. Managing a large number of parallel sessions is terminal-only. No unified diff panel, no session status overview, no Paradev-style workspace UI.
  • Agent Teams still experimental. The coordinated-team primitive is not production-ready after nine months behind a flag.

Which to reach for

Reach for Paradev when: you run multiple agent CLIs (Claude Code plus Codex, Cursor Agent, or Gemini CLI together), you want a visual workspace to manage 20+ parallel sessions without terminal juggling, your team needs cloud execution when machines are off, or you want to migrate from Conductor with minimal workflow change. The MIT desktop is a direct replacement for Conductor's local app. The multi-vendor routing is a capability nothing else in this category matches.

Reach for Claude Code natively when: you are Claude-only and want the simplest path to parallel execution without an additional tool, you have large-scale batch tasks (codebase audits, security sweeps) where dynamic workflows can write the orchestration strategy for you, or you want to stay on first-party primitives as Claude Code continues to ship. No additional subscription needed.

Conductor's exit removes the dominant free parallel runner from the category. Paradev's MIT move positions it as the natural successor for local-only parallel work, while Claude Code's dynamic workflows cover the other end (fully automated large-scale fan-out). For developers who used Conductor primarily as a visual worktree manager for a handful of parallel sessions, Paradev is the most direct replacement. For developers who used Conductor to run a lot of independent tasks they briefed manually, native Claude Code subagents or dynamic workflows cover the same ground without an additional app. For the broader category landscape, the agentic coding patterns that actually scale guide has the full breakdown of which parallel approaches hold up at different task sizes. The Conductor vs Paradev piece from earlier this month has more detail on how those two compared before the acquisition.

The process layer both skip

Paradev manages how many agents run. Claude Code's dynamic workflows automate how agents fan out. Neither answers a different question: what process governs what agents build, in what order, and whether the output is correct before it ships.

That is what defract was built around: a gated lifecycle (story, design, architecture, implementation, review, release) running on Claude Code, with a visual design stage where direction gets locked before implementation starts, and specialist review agents that audit each other before you approve the diff. It is not competing with Paradev on agent volume or with dynamic workflows on orchestration automation. It sits at the stage above both: the process layer that runs before agents start and gates each transition. For teams where the bottleneck is coordination and correctness rather than throughput, why AI coding agents need hard stage boundaries explains the problem it is solving.

defract

a structured lifecycle for your parallel Claude Code agents. free, no caps, no signup.