The parallel-agent workflow stopped being a frontier a few weeks ago. Cursor 3's Agents Window, Windsurf 2.0's free parallel agents, and Anthropic's April 14 Claude Code desktop redesign (multi-session sidebar, per-session worktrees, rebuilt diff viewer, Routines for scheduling) all ship the same core idea: run many agents in isolated worktrees from one surface. If you were still juggling raw terminal tabs last week, upgrade this week.
I've been running four to six parallel sessions every day for the last two months, across and ahead of these releases. Here's what the new tools still don't fix.
1. A session list is not the same as knowing what an agent is doing. A sidebar with a row per session and a status chip beats zsh, zsh 2, zsh 3. It doesn't tell me that session 3 is stuck reconciling three conflicting test fixtures and that I already have notes about those fixtures in another session from last Tuesday. The session is a row. The work is a richer object than a row.
2. Finding the pinging session is better, not solved. "Session 3 needs input" is a real improvement over terminal bells. What's missing is the connective tissue: I want "session 3 (refactor file watcher, linked to tracker #432) is asking whether to keep the old onChange signature" so I can answer in context instead of alt-tabbing into a transcript and reading back.
3. Cross-session diff review is still brutal. Every new tool rebuilt the diff viewer. None of them handle the common case where three parallel agents touch coupled code (file watcher + new IPC handlers + tests for both) and need a single combined review, not three separate ones.
4. Context handoff between agents is still manual. Agent A designs a data model. Agent B writes the migration. Agent C writes tests. Each agent can read the code. What they can't recover is the reasoning from the previous session's transcript, which is where most of the important context lives. Every tool has a transcript, and every transcript is trapped in the session that produced it. It's worse when you mix Claude Code and Codex in the same day (I do — they're better at different things), because now the transcripts aren't even in the same tool. I still end up copy-pasting.
5. Scheduled work needs the whole workspace. Claude Code Routines and Cursor's scheduled agents are here. What's still missing is scheduling that can read and write across the whole workspace: open sessions, tracker items, notes, decisions, yesterday's transcripts. A stateless scheduled script does 60% of that. A scheduled agent with full workspace context does all of it.
The parallel-agent layer is now table stakes. Session management with worktrees is shipped by every serious AI coding tool. The workspace around the sessions (work as a first-class object, cross-session review, shared context that travels, scheduled agents with full workspace access) is still mostly empty, and it has to work across the agents you actually use, not just one vendor's. I'm building into that gap with Nimbalyst, a desktop workspace that runs sessions across Claude Code and Codex in the same project, treats every session as a card on a kanban, keeps transcripts, notes, diagrams, and data models in the same file tree so any agent can read them, and runs scheduled automations that share that context.
If you've hit these same gaps, I'd love to hear how you're solving them. Drop a comment or reply with your setup.
Original article on nimbalyst.com/blog.
Top comments (0)