What changes after chat and copilots?
Think about the work you hand over. In chat, you usually ask for an answer. With a task agent, you assign a defined job. With a persistent agent, you can assign a continuing responsibility.
These categories overlap. Products described as copilots can include agent functions. Task tools can already use applications and schedules. Check the actual capabilities instead of treating the product name as a standard.
| Tool type | What you hand over | Example |
|---|---|---|
| Chat | A question | Explain the differences between seven products. |
| Copilot | Assistance while you work | Help me edit this product comparison. |
| Task agent | A defined job | Research the products and prepare a comparison document. |
| Persistent agent | Work that continues over time | Check the products each week. Update the comparison. Report important changes. |
How does work continue between conversations?
A persistent agent combines several functions. Memory keeps useful context. Application connections give it access to information and actions. Schedules or other triggers start work. A computer environment lets it use software. Permission controls limit what it can do.
These functions differ between products. A cloud agent can continue after you close your laptop. Work that needs your local computer may stop when that computer is unavailable. Memory can also become outdated. Ask the agent to check current sources before important decisions.
Compare seven approaches
This comparison uses provider documentation checked on 30 October 2026. It is not a hands-on reliability test. Dots is an OpenAI product. Muse is a Meta product. The table describes documented differences, not a ranking.
| Product | Main distinction | Access and limits |
|---|---|---|
| OpenAI Dots | Ongoing projects in ChatGPT, with a cloud computer and connected applications. Can delegate tasks to Codex and ChatGPT Work. | The Pro rollout excludes the UK. Business Premium availability depends on region and workspace eligibility. Enterprise beta needs admin enablement. Deeper work has an allowance. |
| Grok Bot | Specialist bots can coordinate work, retain role context and run routines on a persistent cloud computer. | Plan and usage rules depend on the subscription. Bots on one account share files, browser sessions and logins. Separate bots are not security boundaries. |
| Muse — Meta | Personal administration, appointments, shopping and longer goals through the Muse app and WhatsApp. | Meta’s announcement describes a US rollout, free use and subscription options. It does not establish UK availability. Sensitive actions need permission. |
| Claude Duet — Anthropic | Proactive professional assistance, morning briefs and tasks across connected applications. | US research preview through a waitlist for Pro and Max. Background work uses existing plan limits. Checks happen every few hours, not continuously. |
| Gemini Spark / Gemini tasks — Google | Connected applications, recurring tasks and reusable skills. Current help presents tasks alongside chat. | Tasks require a personal account, AI Pro or Ultra, Keep Activity and Personal Intelligence. UK access and work or school accounts are currently excluded. |
| Hermes Agent — Nous Research | Configurable agent with persistent memory, reusable skills, scheduled work and model-provider choice. | You choose where it runs and arrange model access. Skills can be created and revised. This does not prove that the underlying model improves after each task. |
| OpenClaw | Open-source assistant connecting models, devices and messaging channels through a gateway you operate. | You arrange compute and model access. Stored state stays on your hardware, but prompts can reach configured providers. Main-session tools use the host unless sandboxing is configured. |
- OpenAI: Dots plans, eligibility and controls ↗
- Grok Bot: capabilities, shared computer and billing FAQ ↗
- Meta: Introducing Muse — September 2026 ↗
- Anthropic: Claude Duet research preview ↗
- Google: Introducing Gemini Spark — 19 May 2026 ↗
- Google: Current requirements for tasks and workflows ↗
- Nous Research: Hermes Agent documentation ↗
- OpenClaw: Official project, architecture and security notes ↗
Start with availability and the work you need
For a UK business, access is an early filter. Do not assume a US launch is available here. Check the exact plan, account type and workspace controls before you arrange a pilot. A personal subscription may not meet your organisation’s requirements.
Use the comparison to make a shortlist. Consider Dots for continuing ChatGPT projects where your workspace qualifies. Consider Grok Bot for coordinated specialist roles. Look at Hermes or OpenClaw when you need configurable infrastructure and can operate it. Duet, Muse and Gemini remain subject to the access limits above.
These are judgments about documented fit. They do not establish which product will complete your work most reliably. Compare products on the same task before choosing.
Illustrative example: a weekly supplier check
An operations team compares seven software suppliers. Someone currently opens their websites, checks changes and edits a shared document each Friday. A persistent agent could do the first pass and prepare a report for review. This is an illustrative workflow, not a Fewzen client result.
Give the agent a clear instruction:
- Every Friday at 09:00 Europe/London, check the seven approved supplier websites.
- Use official documentation. Record the source URL and the date checked.
- Compare current prices, UK availability and functions with the previous report.
- Update the agreed comparison document. Keep a record of what changed.
- Mark missing or conflicting information. Do not turn an unavailable page into a claim that nothing changed.
- Ask for approval before contacting suppliers, buying anything or changing accounts.
Set the outcome, boundaries and checkpoints
Define a useful result before connecting accounts. For the supplier example, the result is an accurate comparison with evidence. A completed run is not enough if its links are broken or important changes are missing.
Connect only the sources and destination needed for the task. Separate reading information from sending messages or changing records. Set those limits in product controls where available, as well as in the instruction. Name the person who handles exceptions.
Plan how to pause the task, remove access and correct stored information. Self-hosting does not automatically keep all data local. A connected model or messaging service may still process it. Check the actual data path.
Check limits, reliability and total cost
“Always on” does not mean unlimited or constantly watching. Anthropic says Duet checks accounts every few hours and uses plan limits. OpenAI distinguishes read-only proactive research from actions taken under your instructions and controls. These are different operating rules.
An agent can miss a change, misunderstand a page or repeat an action after a failure. Agree how it reports blocked work and uncertain results. Require a record of completed actions so a person can investigate.
Compare the cost of accepted work. Include the subscription, model use, hosting, setup, review and maintenance where applicable. Free software does not mean free operation. Avoid choosing a product from its monthly headline price alone.
Run one small pilot
Start with public information and one output that a person can check. Keep a manual comparison for the same period. Record completed work, missed changes, corrections, review time, elapsed time and total usage cost.
Include an unavailable source and a changed instruction. Check whether the agent explains the problem and asks for help when needed. Expand its responsibility only when the results justify it.
The next step is to write a short task brief: the result, sources, schedule, allowed actions, approval points and exception owner. Then test it with a product you can actually access.