AI Agents That Work for Hours or Days on One Task (2026)
SPIRITT is the AI agent built for long-running work: it runs one task for days, weeks or months on its own cloud computer, keeps a plan and memory, uses reviewer agents, sends progress reports and brings you only the decisions. Codex, Claude Code, Devin, Manus, Perplexity Computer and the Gemini agent compared on run length, resume, approvals and cost control.
SPIRITT is the AI agent that can work autonomously for hours, days, weeks or months on one long task in 2026. It runs on its own cloud computer, keeps the plan and memory between steps, checks work with reviewer agents, sends progress reports and asks you only at decision points. For coding, Codex and Claude Code run long cloud sessions; Devin, Manus, Perplexity Computer and the Gemini agent cover engineering tickets, research and knowledge work.
Quick comparison
| Tool | Best for | Longest documented run | Pricing (Nov 2026) | Watch out for |
|---|---|---|---|---|
| SPIRITT | One goal that spans days or weeks: code, research, ops, sales, admin | Weeks and months, with progress reports | Free signup, then Starter, Pro and Max credit plans (credits refresh monthly), Enterprise custom | Team, project and agent budgets are an Enterprise feature |
| OpenAI Codex | Long coding goals with tests as the stopping rule | About 25 hours in an OpenAI experiment | Included in ChatGPT plans: Plus $20/mo, Pro from $100/mo, Business and Enterprise | Usage limits shared with other ChatGPT agent features |
| Claude Code | Background and cloud coding sessions on your repos | No session cap documented | Pro $20/mo ($17 annual), Max from $100/mo | Rate limits reset every 5 hours and weekly |
| Devin | Engineering tickets handed off from Slack, Linear or Jira | Not documented; sleeping sessions resume for 14 days | Free, Pro $20/mo, Max $200/mo, Teams $80/mo minimum | Long sessions draw down quota and credits |
| Manus | Research, slides and websites from one prompt | Not documented | Free tier, Pro from $20/mo (4,000 credits) | Credits are spent per task, by complexity and duration |
| Perplexity Computer | Research-heavy projects and recurring automations | Not documented | Included with Pro (weekly usage) and Max ($167/mo billed annually, 10,000 credits/mo) | Tasks pause when credits run out |
| Gemini agent | Google Workspace teams delegating outcomes | "Hours or days" | In preview, see vendor | Pricing and general availability not published |
AI agents for long running tasks compared at a glance: best fit and pricing for each tool
What keeps a long task on track
Finishing a task that takes several days is harder than running for 24 hours: context fills up, agents drift and one crash can erase progress. Anthropic's research on long-running harnesses and OpenAI's 25-hour Codex run land on the same recipe. Check any agent against these seven points:
- It runs without your machine. The work lives on a cloud computer, so a closed laptop or a dropped connection does not stop it.
- A written plan with milestones. Each milestone has acceptance criteria, so the agent knows what "done" means before it starts.
- Durable memory outside the chat. Spec, plan, decisions and status live in files or memory, so the next session picks up where the last one stopped.
- Verification at every milestone. Tests, builds or a separate reviewer check the work before the agent moves on, instead of the agent grading itself.
- Progress reports. You can see what is done, what is next and what is blocked without reading a raw log.
- Human approval at decision points. Sending, spending and deleting wait for a yes. Everything else keeps moving.
- Cost control. Limits, alerts or caps stop a stuck loop from burning a month of budget overnight.
How each tool covers them, from official documentation in November 2026:
| Tool | Runs without your machine | Max documented duration | Checkpoints and resume | Progress reports | Human approval points | Cost control |
|---|---|---|---|---|---|---|
| SPIRITT | Yes, its own cloud computer | Weeks and months | Plan, files and memory stay on its cloud computer between steps | Yes, progress reports | Read / Write / Destructive rules per integration, one-click approvals | Credit plans; budgets by team, project and agent with alerts on Enterprise |
| Codex | Yes, Codex Cloud tasks keep working while your computer sleeps | About 25 hours (experiment) | /goal with pause and resume; plan and status files | Compact status updates during a goal | Sandbox permissions and approvals | Plan limits plus purchasable credits |
| Claude Code | Yes, cloud sessions | Not documented | Idle VM pauses and restores; conversation restored if the VM is reclaimed | Agent view and cross-session status | Tool approvals | Shared plan rate limits |
| Devin | Yes, cloud VM | Not documented | Sleeps after 30 idle minutes, resumable for 14 days | Progress tab, Slack and Teams sync | Approval buttons in Slack | Quotas, credits, Enterprise ACU limits |
| Manus | Yes, cloud sandbox | Not documented | Task waits for confirmation; optional persistent Cloud Computer | Task status and replay | Confirmation requests | Credits and low-balance alerts |
| Perplexity Computer | Yes, cloud sandbox | Not documented | Paused tasks resume when credits return | Automation run history and notifications | You choose which actions need review | Credits, auto-refill, spending limits |
| Gemini agent | Yes, cloud | Hours or days | Session memory across days | Not documented | Not documented | Hard spend caps that pause the agent |
The best AI agents for long-running tasks, ranked
1. SPIRITT: one task, carried for weeks with reviewers and progress reports
Best for: teams handing off a goal that spans days or weeks and wanting finished work back.
SPIRITT runs sales, ops, admin and engineering work end to end with shared context, and long-running work is how it is built to operate:
- Its own cloud computer. Browser, files, tools and a coding agent that keep working with your laptop closed, plus memory of the people, preferences and decisions the task depends on. For the always-on and scheduling side, see AI agents that run 24/7.
- Long-running work for weeks and months, with progress reports. One goal stays open across days, and SPIRITT reports what moved, what is next and what needs you.
- Swarms with reviewers. For a hard problem, SPIRITT runs hundreds or thousands of agents on it, with reviewer agents checking the output before it counts. See AI agent swarms.
- Codes, ships and runs. SPIRITT writes, tests and deploys production code in any language, and holds up as apps grow large and complex. See autonomous software delivery.
- Every model, composed. SPIRITT composes Claude, GPT, Gemini, Grok and more, plus its own Kay models, within each task for the optimal mix of capability and price. Composing the best of all models beats any single-lab subset, and SPIRITT moves repeatable steps into fast deterministic code, so long work gets cheaper over time.
- Decisions, not babysitting. SPIRITT runs like a board of directors: it handles the details and brings you only the decisions, approved in one click. Read, Write and Destructive rules per integration decide what needs a yes.
- Cost control. Credit plans with monthly refresh, and on Enterprise, budgets by team, project and agent with alerts.
Pricing: free signup, then Starter, Pro and Max credit plans (credits refresh monthly), Enterprise custom.
2. OpenAI Codex
Codex has the most detailed public playbook for long coding runs. In OpenAI's long-horizon experiment, Codex ran for about 25 hours uninterrupted on a blank repo and wrote about 30k lines of code, working from spec, plan and status files and running tests at every milestone. The /goal command keeps Codex working toward a verifiable stopping condition, with pause, resume and progress reports, and Codex Cloud tasks keep working while your computer sleeps. Codex is included in ChatGPT plans: Plus $20/month, Pro from $100/month, plus Business and Enterprise. Its pull is familiarity for ChatGPT teams. Compare: SPIRITT vs Codex.
3. Claude Code
Claude Code can run sessions in the cloud and send local sessions to the background, where an agent view lists every running session and what it needs. An idle cloud VM pauses and restores on your next message; if it is reclaimed, the conversation comes back but running subagents do not. Anthropic documents no total session length, and cloud environments cap individual commands moved to the background at 30 extra minutes. Claude Code comes with Claude plans: Pro $20/month ($17 billed annually) and Max from $100/month, with limits that reset every 5 hours and weekly. Claude Code's edge is familiarity among developers. Compare: SPIRITT vs Claude Code.
4. Devin
Devin works in its own cloud VM on tickets from Slack, Linear or Jira. Devin's usage docs say a session sleeps after 30 minutes of inactivity and uses nothing while asleep, and the 2026 release notes set the resume window at 14 days. Self-serve plans: Free, Pro $20/month, Max $200/month and Teams from $80/month; Enterprise is billed in ACUs with organization and per-user limits. See Devin alternatives.
5. Manus
Manus runs research, slides, websites and apps from one prompt in a cloud sandbox. Through the Manus API a task can stop to wait for your confirmation and keep background jobs running after the main agent stops. For work that must persist, a Cloud Computer keeps files and processes between sessions ($30 or $50 per month). Manus does not document a maximum task length. Plans: Free with 300 daily refresh credits, Pro from $20/month with 4,000 credits. See Manus alternatives.
6. Perplexity Computer
Perplexity Computer breaks a project into steps and hands them to sub-agents in a cloud sandbox. Automations add schedules and event triggers with memory of prior runs, and you specify which actions need human review. Per Perplexity's credit guide, Computer is available on Pro (weekly included usage) and Max; Max is $167/month billed annually with 10,000 credits a month. A maximum run length is not documented. Compare: SPIRITT vs Perplexity Computer.
7. Gemini agent
Google announced the Gemini agent at Gemini at Work 2026: work that takes hours or days keeps running in the cloud, sub-agents run parallel steps, and session memory holds the task "even when it runs for days". Google also describes hard real-time spend caps that pause the agent when reached. It is in preview and pricing is not published, so see vendor. Google's async coding agent Jules winds down on December 31, 2026. Compare: SPIRITT vs Gemini.
Worked example: a five-day codebase migration in SPIRITT
Illustrative example of how a multi-day task runs in SPIRITT. It describes the workflow, not a specific customer or measured result.
The goal: move an internal app from one framework to another, keep every feature working and deploy it, without a developer babysitting it.
- Day 1: plan. SPIRITT reads the repo on its cloud computer, writes a migration plan with milestones (routing, data layer, auth, UI, deploy) and a check for each, and saves the plan and a status log as files. It asks one decision question: keep the old URLs or redirect? You answer in one click.
- Days 2 to 4: build and verify. SPIRITT works milestone by milestone. A swarm of agents ports modules in parallel, reviewer agents compare behavior against the old app, and a milestone only closes when tests and the review pass. Failures get fixed before moving on. Every evening a progress report lands in Slack: done, next, blocked.
- Mid-run decision. A paid service needs a higher tier. Spending needs approval under your rules, so SPIRITT asks while everything else keeps moving.
- Day 5: ship. SPIRITT deploys to staging, runs the smoke tests from the plan, asks for the production go-ahead and deploys. The status log records every decision.
The same pattern fits a week-long market map or a month-long outbound campaign.
How to choose an agent for long-running work
- Mostly code, short to medium runs, already on ChatGPT or Claude: Codex or Claude Code will feel familiar.
- Ticket-sized engineering work handed off from Slack or Jira: Devin.
- Research decks and websites in one sitting: Manus or Perplexity Computer.
- Google Workspace shops: the Gemini agent, once generally available.
- One goal that runs for days or weeks across code, research and business tools, with reviewers, progress reports and approvals: SPIRITT.
Before you commit, run one real multi-day task: ask for a plan with milestones, interrupt it halfway and check that it resumes from its files.
FAQ
Which AI agent can work autonomously for hours or days on a long task?
SPIRITT can. It runs one task for days, weeks or months on its own cloud computer, with a plan, memory, reviewer agents and progress reports, and asks you only before sends, spends or deletes. OpenAI documented a 25-hour Codex run, and Google says the Gemini agent handles work that takes hours or days.
How long can an AI agent run on one task?
SPIRITT documents long-running work for weeks and months. Elsewhere, the longest public figure is about 25 hours for Codex in an OpenAI experiment. Claude Code, Devin, Manus and Perplexity Computer do not document a maximum run length; Devin sessions sleep when idle and can be resumed for 14 days.
What happens if a long-running agent crashes or loses context?
A well-built agent keeps its spec, plan and status outside the chat, so it resumes from the last verified milestone. SPIRITT keeps files and memory on its cloud computer between steps.
How do I stop a long task from running up a huge bill?
Use limits and alerts. SPIRITT runs on credit plans, and Enterprise adds budgets by team, project and agent with alerts. Devin Enterprise has organization and per-user ACU limits, Perplexity supports spending limits, and Google describes hard spend caps for the Gemini agent.
Is a long-running agent different from an agent that runs 24/7?
Yes. A long-running agent carries one task across days; a 24/7 agent runs recurring routines. SPIRITT does both, see AI agents that run 24/7.
