Current agent options to know / Checked October 22, 2026
Start with the work, the context and the amount of control you need. Then compare tools on the same small task.
This is a source-checked capability guide, not a popularity ranking or a hands-on performance benchmark. Access changes by plan, region, device and rollout. Follow the official links before buying.
| Layer | Plain-language meaning | What you decide |
|---|---|---|
| LLM / model | The engine that interprets context and generates content or tool requests. | Choose capability, speed and cost against your test cases. |
| Agent | A system that pursues a goal by choosing steps, using available tools and checking results. | Define the goal, authority, stopping point and owner. |
| Harness / runtime | The surrounding software that manages the loop, tools, state, files, permissions and execution. | Check where it runs, what it can reach and which controls are enforced. |
| Workflow | The steps and handoffs that turn an input into a useful result. Some steps can be fixed; others can use an agent. | Choose the simplest dependable process. |
| Skill / plugin / connector | A skill packages reusable instructions or resources. Plugins and connectors add specific capabilities; terminology varies by vendor. | Check what is installed, connected and allowed. |
| MCP | Model Context Protocol: a standard way compatible applications discover and use tools and context from servers. | Verify the server, permissions and data source. MCP itself does not certify trust. |
| Retrieval / RAG / memory | Retrieval brings selected source material into a task. RAG combines retrieval with generation. Memory retains selected context across work. | Set freshness, access, source precedence and deletion rules. |
Further reading: agent and workflow patterns, harnesses and evaluation, MCP architecture, and skills and plugins.
Task-oriented workspace
Use it for multi-step work with files and approved tools: a sourced brief, analysis, spreadsheet, deck or recurring workflow. It emphasizes the finished deliverable and can use plugins.
Work and Codex overlap. Codex exposes more development details such as diffs and PR review; it is not restricted to coding. Local-resource steps still need a connected computer.
Official Work guide → · Work and Codex →Task-oriented workspace
Useful for research, files, reports, spreadsheets and other multi-step knowledge work. Anthropic is rolling Chat and Cowork into one conversation, beginning with Pro and Max.
Some learners will still see separate modes. Cloud tasks can continue after leaving; tasks using local apps or files require the desktop app to remain available.
Official transition guide →Development and technical work
Choose a development workflow when the task centers on code, files, commands, tests, repositories or a technical automation. Both can support tools and reusable project instructions.
Ask for a plan, inspect changes, run meaningful tests and verify the deployed behavior. Generated code and a successful local run do not establish production readiness.
Codex documentation → · Claude Code →Visual work
Create and iterate on designs, prototypes, one-pagers and visual artifacts. The current experience is integrated with conversations, Artifacts and Claude Code.
Anthropic says the standalone Design site closes December 14, 2026. Learn the integrated flow. Check behavior, readability, brand fit and accessibility separately from appearance.
Official Design guide →These products can continue work between messages. That makes context, approvals, budgets and a clear pause mechanism part of the job definition.
Dots have a cloud computer, connected apps and ongoing project context. They can use Codex and other tools and bring work or decisions back for review.
Try: keeping a project brief current and preparing the next decision. Check your account’s access, custom rules and available connections.
Official dots overview →Meta documents a personal agent with its own cloud VM, browser and code tools, background work and subagents. Its architecture separates the agent from credential and permission services.
Try: a bounded research or project responsibility. Review the evidence and approval controls; the vendor explicitly acknowledges remaining errors and attacks through retrieved data.
Meta’s Muse architecture →Announced October 21, 2026, Duet is a proactive agent that uses connected-app context, prepares briefs and keeps working on tasks. The research preview starts with a waitlist for US Pro and Max subscribers.
Try when available: client-meeting preparation and follow-up drafts. Background use shares plan limits. Preview access is not a prerequisite for this course.
Official Duet announcement →Named Bots can retain role context, use a persistent cloud computer, coordinate work and run skills or routines. Work can continue while your laptop is closed.
Try: a research Bot handing a source brief to a reporting Bot. All Bots on an account share files, browser sessions and logins; separate names do not isolate data.
Official Grok Bot FAQ →Use explicit steps for a predictable business process, with bounded agent work inside it. Add validation, retries, a human review tool and a final write step.
Useful when an event should reliably move through defined checks. You own credentials, error handling and operations.
n8n AI Agent documentation →Evaluate the agent builder and governance tools that fit your organization’s existing Microsoft or Google environment. Enterprise access, triggers and administration differ from consumer chat plans.
Copilot Studio → · Gemini Enterprise →Build a custom service when the workflow needs application code or controls a ready-made product cannot provide. Managed runtimes, SDKs and custom loops assign different responsibilities to you.
OpenAI agents → · Claude Agent SDK →Other options to investigate for a specific task include Manus for cloud-browser work, Genspark for multi-step artifact work and Lindy for connected assistant workflows. Compare the controls on the exact surface you plan to use.
Use a coordinator and specialists when the work splits into independently useful pieces: a research scout, an analyst and a reviewer can return structured outputs to one owner. Give each a narrow question, approved tools, an output format, a budget and a stopping point.
Use sequential handoffs when later work depends on earlier evidence. Avoid two agents editing the same file without ownership rules. More agents add communication, usage and review; they are useful only when they improve the outcome. A reviewer agent can miss the same mistake as the maker.
OpenAI subagents · Claude Code subagents
Start with the available default in an approved product and evaluate it on your task. Compare capability, speed, context, tool support and total cost. Escalate when a specific failure justifies a stronger model or more reasoning effort.
Current OpenAI guidance identifies GPT-6.1 Sol for complex agentic work, GPT-6 Luna for focused repeatable volume and GPT-6 Astra for demanding work. Anthropic’s current guide lists Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5. These are dated examples, not a cross-vendor ranking.
Current OpenAI model guide · Current Claude model guide
Count all model calls, subagents and retries, then add tools, compute and human preparation, review and rework. Use concise handoffs, retrieve relevant context, cap steps and retries, and cache stable context where supported. Recheck quality after each cost-saving change.
The planner’s cost example uses your rates. Subscription allowances and API billing are different. See official usage and observability guidance.
Observed now: persistent agents, cloud execution, connected tools and visual creation are converging with conversational interfaces. The product transitions above are documented examples.
Near-term expectation, not a promise: more workflows will combine background execution, organization policies and human review. Skills and context may become easier to reuse, while identity and permission controls remain important.
Longer-term scenarios: people may supervise more portfolios of agent-assisted work, or adoption may advance unevenly because of trust, cost and integration limits. No specific timeline or autonomous-business outcome is guaranteed. Learn to specify, evaluate and own work across either scenario.