Which long-running coding agent is available in production today?

All three have production-available variants, but with very different maturity. Cursor Cloud Composer has been generally available since Cursor 3 (late 2025) and saw a major upgrade with Cursor 4's auto-router in May 2026 — most-mature consumer-facing option. Devin Desktop (Cognition's rebranded long-running agent product, evolved from the original Devin SaaS) has been GA since early 2026 and is positioned at enterprise teams. Codex+Ona is the newest entrant — OpenAI announced the Ona acquisition June 11, 2026, with integration subject to regulatory approval; the existing Ona standalone product is GA today and existing Codex CLI is GA, but the integrated Codex+Ona experience is targeted for Q3-Q4 2026.

Which one runs inside my own cloud (VPC) vs the vendor's cloud?

Codex+Ona is the only one of the three that runs agents inside the customer's own VPC by default. That's the entire pitch — Ona's platform was built around customer-controlled execution environments with kernel-level policy enforcement, and OpenAI bought Ona specifically for that capability. Cursor Cloud Composer runs in Cursor's cloud infrastructure with optional VPC peering for enterprise customers. Devin Desktop runs in Cognition's managed cloud by default with private deployment options for enterprise. If regulatory or data-residency requirements force customer-VPC execution, Codex+Ona (post-integration) is the strongest fit; today, the existing Ona standalone product is the only fully customer-VPC option among the three.

How do their cost models differ?

Three different shapes. Cursor Cloud Composer: subscription plus metered compute — Cursor Ultra at $200/month includes a generous compute allowance, with overages metered. Costs are predictable; vendor manages everything. Devin Desktop: enterprise-only seat pricing, typically $500-$2000 per engineer per month depending on usage tier, with compute included in seat price. Most predictable but most expensive. Codex+Ona: customer pays cloud bill (AWS/GCP/Azure) for agent execution + OpenAI model API costs + (post-integration) OpenAI orchestration fees. Lowest marginal cost at scale if you have committed cloud spend, but two-vendor bills and harder to forecast. For a heavy engineering team, expect $300-$2000 per engineer per month across all three; differences come in cost predictability and where the money goes.

Which is best for true multi-day autonomous agents?

Honestly, none of them yet, despite the marketing. Mid-2026 reality: 'agents that run for days' is mostly aspirational. The practical state of the art is agents that survive overnight (6-12 hours) without supervision and produce a coherent pull request by morning. All three achieve this for well-scoped tasks. Beyond ~12-24 hours, context-window limits, drift, and review-velocity bottlenecks dominate. If you genuinely need multi-day agent runs, plan on checkpointing strategies (resumable agents that summarize and restart), human checkpoints at logical boundaries (24-hour reviews), and small experiments before committing major migrations. The marketing framing of all three vendors is ahead of reality; treat 'days' as 'overnight, multiple times in a row with human checkpoints.'

Quick Answer

Codex+Ona vs Cursor Cloud Composer vs Devin Desktop: Long-Running Coding Agents (June 2026)

Published: June 15, 2026

Codex+Ona vs Cursor Cloud Composer vs Devin Desktop: Long-Running Coding Agents (June 2026)

Three approaches to coding agents that run for hours or days without your laptop being on. OpenAI’s Codex+Ona (announced June 11, 2026, GA Q3-Q4) runs agents in your own VPC. Cursor Cloud Composer runs in Cursor’s cloud and ships in Cursor 4. Devin Desktop runs in Cognition’s cloud and targets enterprise. This page compares the three on architecture, cost, availability, and what each is genuinely good at.

Last verified: June 15, 2026.

TL;DR

Codex+Ona — Customer VPC, deepest enterprise security story, GA Q3-Q4 2026 (regulatory pending).
Cursor Cloud Composer — Available now, broadest consumer adoption, IDE-native.
Devin Desktop — Available now, enterprise-only, highest-touch managed experience.
None of them yet deliver “agents that run for days” in production. Treat as “overnight + checkpoints.”

Side-by-side

Dimension	Codex+Ona (post-integration)	Cursor Cloud Composer	Devin Desktop
Vendor	OpenAI + Ona (acquired June 2026)	Cursor (Anysphere)	Cognition
Execution location	Customer VPC	Cursor cloud	Cognition cloud
Availability	GA Q3-Q4 2026 (regulatory pending)	GA today	GA today
Default model	GPT-5.5 / Codex-class (then GPT-5.6)	Auto-router (Fable 5, Opus 4.8, GPT-5.5, etc.)	Claude Opus 4.8 / Fable 5
Trust boundary	Customer-controlled	Cursor-managed	Cognition-managed
Agent duration target	Hours to days	Minutes to hours	Hours to days
Pricing model	Cloud bill + OpenAI API + orchestration	Subscription + metered	Enterprise seat
Typical price (heavy use)	$300-$1500/dev/mo	$200-$500/dev/mo	$500-$2000/dev/mo
IDE integration	CLI + future Codex integrations	Native Cursor IDE	Web UI + CLI
Audit / compliance	Kernel-level, customer-controlled	Cursor enterprise tier	Cognition enterprise tier
Best for	Regulated industries, customer-cloud requirements	Solo devs and small teams who live in Cursor	Enterprise engineering orgs

When each one wins

Codex+Ona wins when

Regulated industry, customer-VPC required. Banks, hospitals, defense, government — the kernel-level customer-controlled execution model is genuinely differentiating. Ona’s existing customer list (BNY, Pearson, GSR, Vanta, EquipmentShare, Hargreaves Lansdown) signals strong fit.
You’re on the OpenAI stack and have AWS/GCP commitments. Codex models run on OpenAI API; agent execution runs on your cloud. Double up on existing vendor relationships.
You can wait until Q3-Q4 2026. The integration is subject to regulatory approval and broader GA isn’t imminent.

The current state (June 15, 2026): Ona standalone is GA and works today; Codex CLI is GA; the integrated Codex+Ona experience targeted at OpenAI customers is on the roadmap, not yet live.

Cursor Cloud Composer wins when

Cursor is your IDE. Cloud Composer integrates natively with the Cursor editor — same chat surface, same agent affordances, just hands work off to the cloud when you close your laptop.
You want a single subscription that covers IDE + cloud agent. Cursor Pro/Ultra includes Cloud Composer access; you don’t manage cloud bills.
Auto-router fits your workflow. Cursor 4’s auto-router picks Fable 5 / Opus 4.8 / GPT-5.5 / Gemini 3.5 Pro per task. Best for teams that don’t want to think about model selection.
Volume is low to moderate. $200/mo Ultra covers a heavy individual; you pay overages above the included quota.

Devin Desktop wins when

You’re an enterprise engineering org and want a managed long-running agent. Cognition’s white-glove support and tuning for long-running workflows is what you’re buying.
Your team prefers a web UI + CLI over IDE-native integration.
You’re already a Devin customer. Existing deployments roll forward; Devin Desktop is the new flagship surface.

The Codex+Ona timeline

OpenAI announced the Ona acquisition on June 11, 2026. Realistic timeline:

Now → Q3 2026: Regulatory review (US antitrust, EU merger control). Ona continues operating independently; existing customers unaffected.
Q3 2026: Approval expected (assuming no second-request). OpenAI begins integration engineering.
Q3-Q4 2026: Closed-beta integrated Codex+Ona to OpenAI enterprise customers.
Q4 2026 - Q1 2027: Broader GA rollout.
2027+: Codex+Ona becomes the default Codex enterprise experience; Codex CLI continues for individual developers.

If you need long-running agents in production today, Codex+Ona is not the answer yet. Cursor Cloud Composer or Devin Desktop are the only two GA options among this set.

The “agents that run for days” reality check

All three vendors lean into multi-day autonomous agent messaging. The honest mid-2026 state of the art:

What actually works:

Overnight runs (6-12 hours) on well-scoped tasks: GA on all three.
Multi-day runs with explicit checkpointing and human review at logical boundaries: feasible, requires careful task decomposition.

What doesn’t reliably work yet:

True 24+ hour autonomous runs with no checkpoints: 10-30% failure rate across vendors. Context drift, infinite loops, missed acceptance criteria are common.
“Set it and forget it for a week” workflows: aspirational. Even Ona’s best-publicized customer success (a top-100 global company doing 4x productivity gain on Python repo modernization) involves human review at frequent intervals.

Three reasons multi-day autonomy is hard:

Context drift. Agents accumulate context until they hit window limits; summarization strategies are still maturing.
Review velocity is the bottleneck. Agents that produce PRs need humans to review. Most engineering orgs review at 10-20 PRs per developer per week; that’s the real ceiling.
Cost compounds. A multi-day agent on Fable 5 or GPT-5.5 can burn $20-$100 in pure model cost; multiply by drift-and-restart cycles and the math gets ugly fast.

The practical recommendation: treat all three vendors as offering “overnight autonomy” and design workflows around that.

Decision flow

Question 1: Do you need agents inside your own VPC?
  Yes → Codex+Ona post-GA, or Ona standalone today.
  No  → Continue.

Question 2: Is Cursor already your IDE?
  Yes → Cursor Cloud Composer is the lowest-friction choice.
  No  → Continue.

Question 3: Are you an enterprise eng org wanting managed long-running agents?
  Yes → Devin Desktop.
  No  → Continue.

Question 4: Do you want a single $200/mo subscription that covers everything?
  Yes → Cursor Cloud Composer.
  No  → Codex CLI today + Codex+Ona Q3-Q4.

Cost model deep dive

Cursor Cloud Composer:

Cursor Ultra at $200/mo per user; Cloud Composer included with generous compute allowance.
Overages metered at published per-task or per-minute rates.
Single vendor bill. Most predictable for solo and small team budgeting.

Devin Desktop:

Enterprise seat pricing, negotiated. Public figures suggest $500-$2000 per engineer per month.
Compute included in seat price.
White-glove support and tuning included at higher tiers.

Codex+Ona (post-integration projection):

AWS/GCP/Azure compute bill for agent execution (variable, but reduces with committed cloud spend).
OpenAI Codex API costs (per-token).
OpenAI orchestration fee (TBD).
Two-vendor bills, harder to forecast but lowest marginal cost at large scale with cloud commitments.

For a heavy team of 20 engineers running long-running agents most weekdays, rough monthly cost estimates:

Cursor Cloud Composer Ultra: $4,000-$10,000
Codex+Ona: $6,000-$30,000 (cloud bill drives variance)
Devin Desktop: $10,000-$40,000

These are rough. Real numbers depend on workload, but the cost ordering is roughly Cursor < Codex+Ona at large scale < Devin Desktop.

Other options worth considering

This page focuses on three vendors but the space is broader. Worth evaluating alongside:

Claude Code background sub-agents — Anthropic-cloud execution, available now, deepest reasoning on Fable 5 / Opus 4.8. See Codex+Ona Cloud Agents vs Claude Code Background Tasks.
GitHub Copilot Workspace — GitHub’s long-running agent tied to GitHub repos and PRs. Tight GitHub integration; weaker than the three above on hard reasoning.
Windsurf Cascade Cloud — Codeium’s cloud agent in Windsurf IDE. Codeium-Cursor-Cline rivalry continues.
Aider + open models on owned infrastructure — DIY path. Cheapest at scale if you have the engineering capacity to operate.

What to watch next 60 days

Codex+Ona regulatory milestones — first procedural filings will signal timeline confidence.
Cursor Cloud Composer + Fable 5 — Cursor’s auto-router has been tuning Fable 5 weights since June 9; watch for benchmark updates.
Devin Desktop pricing changes — Cognition has been adjusting tiers; expect refreshed pricing pages mid-summer.
Anthropic response — Claude Code background sub-agents could gain customer-VPC option as Anthropic responds to Codex+Ona’s positioning.

Codex+Ona integration is subject to regulatory approval; timeline estimates reflect current public guidance and may slip.

Codex+Ona vs Cursor Cloud Composer vs Devin Desktop: Long-Running Coding Agents (June 2026)

TL;DR

Side-by-side

When each one wins

Codex+Ona wins when

Cursor Cloud Composer wins when

Devin Desktop wins when

The Codex+Ona timeline

The “agents that run for days” reality check

Decision flow

Cost model deep dive

Other options worth considering

What to watch next 60 days

Related reading