Future Teams: How Humans and AI Agents Will Work Together

For most of the past three years, working with AI meant one person, one chat window and one task. You asked, it answered, and the conversation was private.

That model ended in 2026. Between June and September, every frontier lab shipped some version of the same idea: an AI agent that has its own computer, its own identity, keeps working when you close your laptop, and sits inside the channels where your team already works. The question for leaders is no longer whether AI can do the work. It is whether your people know how to work alongside it.

This article looks at what has shipped, what people building at the frontier are saying, and four scenarios for how human and agent teams could operate in small and large organisations over the next few years.

---

What the AI Frontier shipped in 2026

Claude Tag: one agent per channel, visible to everyone

Anthropic launched Claude Tag on 23 June 2026, replacing its earlier Claude in Slack app with a persistent agent that lives in team channels, has its own identity and operates under per channel permissions [1]. The central design choice is that it is shared. Within a channel there is a single Claude that everyone can see, delegate to and pick up from where a colleague left off [2].

Slack's general manager, Rob Seaman, summarised the shift by saying that instead of a private exchange, Claude Tag "shows up in the open" [3]. Anthropic paired the launch with its "agent identity" access model: the agent does not act as the user who tagged it, but under service accounts that administrators provision and scope [4].

Alongside the product, Anthropic published the lessons from several months of running human and agent teams internally [5]. They are worth reading in full, but the four headlines are: work in public and give agents broad context; give every human and agent a defined role and the right tools; have humans set a north star so agents can be proactive; and expand agent autonomy deliberately as trust is earned. One practical detail stands out: teams asked their agents to compile a weekly "lessons and missteps" report so the agents would track and avoid their own mistakes [6].

Grok Bot: a team with a chief of staff

xAI opened the beta of Grok Bot on 11 August 2026. Each bot is an always on agent with a cloud computer that signs into your existing tools and completes multi step work [7]. xAI's internal examples are telling: sales bots updating the CRM and drafting follow ups, operations bots processing invoices, and engineering bots reproducing UI bugs and handing fixes to debugging bots [8].

On 28 September, Elon Musk posted a short announcement on X: "Now @Bot works as a team." [9] Bots can now message each other, share context and hand off tasks, with xAI framing the arrangement as a small team led by a chief of staff rather than a stack of automations [10]. The trade off is architecture. Grok Bot's agents share one virtual machine, which the product itself described as a real "blast radius" [11].

OpenAI Dots: always on, and eventually in teams

At DevDay on 29 September, Sam Altman introduced Dots, describing them as "the real-deal AI that we always imagined for ourselves." [12] Each Dot runs on GPT 6 Astra with its own isolated cloud computer, works across ChatGPT, Slack and Microsoft Teams, and is designed to look for work before anyone asks [13]. OpenAI also introduced ChatGPT Space, Pages and collaborative slides so that teammates and agents can edit the same artefacts at the same time [14].

Two details matter for collaboration. First, OpenAI is piloting Specialist Dots with their own organisational identities and credentials, and is working with Microsoft so that IT teams can govern them through Agent 365 [15]. Second, the "team of Dots" is still mostly a promise: users start with a single main agent, with multi agent teams to follow [16]. One early reviewer found that a colleague's Dot, once added to Slack, would only respond to its owner, which made team use difficult [17].

Meta Muse: reliability as the product

Meta's Muse, launched on 8 September, is a consumer agent rather than a workplace one, but it adds a useful design pattern: a second agent called Sentinel sits alongside Muse as a control layer [18]. Meta also gated the launch on more than 100 specific behaviours, each with its own evaluation and threshold [19]. Expect that pattern, one agent doing and another checking, to become standard inside teams.

---

What the frontier is saying

The commentary on X and in founder interviews converges on a few ideas.

Agents will vastly outnumber people inside software. Box CEO Aaron Levie wrote on X in June that "Agents will use software 100X more than people." [20] His point is that systems of record do not disappear; they need controls for identity, approval, audit and rollback, because agents become their heaviest users.

Multiplayer is the unlock. Writing about Claude Tag, one industry commentator argued that the move from single player to multiplayer AI makes "long-running, team-based work possible." [21] Anthropic's developer account put the distinction simply: Claude Code is solo and synchronous, while Claude Tag is multiplayer, asynchronous and proactive [22].

Attention is the new constraint. Altman said in a DevDay interview that he wants his Dot to ping him more often, while acknowledging the risk of companies growth hacking agent notifications [23]. When agents never sleep, the scarce resource is the human judgment they need to escalate to.

Coordination and safety pull against each other. One analysis of the three launches put it plainly: "Coordination and isolation pull in opposite directions." [24] OpenAI isolates each Dot, which contains risk but makes coordination harder. xAI shares one environment, which makes coordination easy but widens the impact of any mistake. Anthropic scopes access per channel. None of these is settled.

---

Synthesis: the technology is ahead of the teamwork

Put these launches side by side and five shifts become clear.

1. From private to shared. The agent is no longer your assistant. It is the team's colleague, and its work is visible.
2. From borrowed identity to its own identity. Agents are getting their own credentials, roles and audit trails, which means they can be governed like staff rather than like personal tools.
3. From reactive to proactive. Agents now look for work. That only helps if humans have defined what good work looks like.
4. From one agent to rosters. Teams will hold several specialised agents, which raises the same questions any manager faces: who does what, who checks whom, and who escalates to whom.
5. From capability to practice. The binding constraint is no longer model quality. It is how well humans delegate, review and decide.

The research supports that last point. A May 2026 paper found that human and AI teams fail to outperform their best individual member in 70% of studies [25]. A September study reported that when people work with an LLM, they capture only about half of the model's standalone accuracy gains, because they defer inconsistently and their confidence becomes poorly calibrated [26]. Meanwhile, Capgemini research suggests that more than 60% of organisations expect to run human and agent teams within twelve months [27].

In other words, the agents are arriving faster than the skills to work with them. Anthropic names the failure mode that fills the gap: everyone quietly running their own private fleet of AI, duplicating work and fragmenting context [6].

---

Four scenarios for the next few years

These are not predictions. They are plausible extrapolations from what has already shipped, intended to help leaders think about design choices now.

Scenario 1: The five person firm with a twenty agent roster (small organisation)

A boutique leadership development consultancy has five people and around twenty agents. The roster is a shared document listing every member, human and agent, with their role, owner, tools and level of autonomy. The founder writes the quarterly north star: grow programme renewals with mid sized professional services firms. Overnight, a research agent scans client news and job postings for signals; a proposal agent drafts tailored programme outlines; a verifier agent checks each draft against the firm's methodology and pricing rules. By morning, the humans review three proposals rather than writing them.

The humans spend their time where agents cannot substitute: facilitating sessions, holding difficult conversations with executives and deciding which clients to prioritise. The firm competes with consultancies ten times its size, not because its people work harder, but because its people are skilled at directing, reviewing and improving agent work. Altman has argued that AI could enable many more one person and small team companies [28]. This is what that looks like when it works.

Scenario 2: The solo operator who becomes a manager on day one (small organisation)

A sole trader coach starts a business and, from the first week, is effectively a manager of six agents covering bookings, invoicing, content, client follow up and session notes. She has never managed anyone. Her early weeks are chaotic: agents send follow ups she has not approved, and two agents duplicate the same newsletter. The fix is not better technology. It is management basics: clear roles, a written escalation rule for anything involving money or a client relationship, and a weekly review of agent mistakes.

The lesson for small organisations: delegation, feedback and role clarity, skills once learned only by people who managed people, are now required from the first day of running a business.

Scenario 3: The bank that hires agents through HR (large organisation)

A regional bank treats agents as part of its workforce. Every agent has an identity issued through IT, a named human owner, a defined scope and an audit trail, much like the agent identity and Agent 365 approaches described above. A new function, part HR and part IT, owns the human and agent roster, which Capgemini's research anticipates in its suggestion of roles such as a VP of Human Agent Resources [27].

Middle managers in this bank change role. They no longer supervise task completion. They design rosters, write north stars for their teams, decide which decisions agents may make alone, and review the weekly lessons and missteps reports. Performance reviews measure joint human and agent outcomes rather than individual output. The managers who struggle are the ones who were promoted for doing the work themselves; the ones who thrive are the coaches and system designers.

Scenario 4: The enterprise with ten thousand shadow fleets (large organisation, cautionary)

A global company buys personal agent licences for every employee but changes nothing about how teams work. Within a year, each person runs their own private set of agents. Three departments produce three different versions of the same market analysis. A sales agent commits to a delivery date that an operations agent knows is impossible, because the two never share a channel. Leadership concludes that "AI is not delivering," when the real problem is that the organisation automated individuals rather than teams.

This is the default path for any organisation that treats agents as a software rollout rather than a change in how people collaborate. It is also the most likely path, given that the research shows most human and AI pairings underperform today [25].

---

What leaders should do now

The difference between Scenario 3 and Scenario 4 is not budget or vendor choice. It is capability. Five practices separate teams that capture the value from those that do not:

1. Work in public. Move agent work into shared channels and shared artefacts so context compounds rather than fragments.
2. Keep a roster. Write down every human and agent on the team, what each owns, and what each may decide alone.
3. Set the north star. Proactive agents need a goal that humans have debated and documented. Without one, proactivity becomes noise.
4. Pair every doer with a checker. Whether it is a verifier agent or a human reviewer, nothing consequential should be marked complete by the agent that produced it.
5. Grow autonomy deliberately. Start with humans reviewing every decision, then expand scope as trust is earned and mistakes are logged and learned from.

None of these are technology problems. They are leadership and learning problems: how to delegate, how to give clear direction, how to calibrate trust and how to build a team culture where work is visible. These have always been the core of good management. What has changed is that every employee now needs them, not only those with direct reports.

That is why we believe the future of collaboration is a learning and development challenge first. At Paldar, we are building tools that help L&D providers and internal L&D teams develop exactly these capabilities, so their people become the humans that make human and agent teams work. If that is a problem you are working on, we would like to hear from you at paldar.com.

---

References

1. Bestblogs, "Claude Tag: Anthropic Brings a Multiplayer AI Teammate Into Slack." https://www.bestblogs.dev/en/explore/topics/anthropic-tag-claude-launch
2. eesel AI, "What is Claude Tag? Anthropic's @Claude AI teammate for Slack." https://www.eesel.ai/blog/claude-tag
3. The Star / Reuters, "Anthropic launches Claude Tag in Slack with plans for wider rollout," 24 June 2026. https://www.thestar.com.my/tech/tech-news/2026/06/24/anthropic-launches-claude-tag-in-research-preview-for-slack-users
4. Mervin Praison, "Claude Tag Explained: Anthropic's Multiplayer AI Teammate for Slack (Beta)." https://mer.vin/2026/06/claude-tag-explained-anthropics-multiplayer-ai-teammate-for-slack-beta/
5. Anthropic, "Building effective human-agent teams," 24 June 2026. https://claude.com/blog/building-effective-human-agent-teams
6. Design Bootcamp (Medium), "Building effective human-agent teams with Claude." https://medium.com/design-bootcamp/building-effective-human-agent-teams-with-claude-9f0025eba1be
7. Build Fast with AI, "Grok Bot review." https://www.buildfastwithai.com/blogs/grok-bot-review
8. KuCoin News, "xAI Launches Grok Bot: An AI Agent with Its Own Computer for Cross-App Task Automation." https://www.kucoin.com/news/flash/xai-launches-grok-bot-ai-agent-with-its-own-computer-for-cross-app-task-automation
9. Basenor, "Grok Bot Now Works as a Team: What That Actually Means" (embeds @elonmusk post, 28 September 2026). https://www.basenor.com/blogs/news/grok-bot-now-works-as-a-team-what-that-actually-means
10. Unite.AI, "xAI Launches Grok Bot: Always-On AI Teammates With Their Own Cloud Computers." https://www.unite.ai/xai-launches-grok-bot-always-on-ai-teammates-with-their-own-cloud-computers/
11. Reworked, "xAI Launches Grok Bot AI Agents in Beta." https://www.reworked.co/collaboration-productivity/xai-launches-grok-bot-ai-agents-in-beta/
12. VKTR, "OpenAI Introduces Always-On Dots Agents," 30 September 2026. https://www.vktr.com/digital-workplace/openai-introduces-alwayson-dots-agents/
13. Mind the Product, "Everything you need to know about Dots, OpenAI's new agent," 2 October 2026. https://www.mindtheproduct.com/everything-you-need-to-know-about-open-a-is-dots/
14. Startup Researcher, "OpenAI Launches Dots Agents, ChatGPT Space, GPT-6.1 Sol, and Ultrafast." https://www.startupresearcher.com/news/openai-launches-dots-agents-chatgpt-space-gpt-6-1-sol-and-ultrafast
15. ProPakistani, "OpenAI Launches Personal AI Agents That Work 24/7." https://propakistani.pk/2026/09/30/openai-launches-personal-ai-agents-that-work-24-7/
16. Letem světem Applem, "OpenAI presented Dots." https://www.letemsvetemapplem.eu/en/2026/09/30/openai-predstavilo-dots-novi-ai-agenti-mohou-pracovat-24-7-a-propoji-se-s-vice-nez-4000-aplikacemi/
17. efficient.app, "Dots review." https://efficient.app/apps/dots
18. Build Fast with AI, "Muse Personal AI Agent Review." https://www.buildfastwithai.com/blogs/muse-personal-ai-agent-review
19. WisdomTree, "From Demo to Distribution: What Meta's Muse Tells Us About Agents at Scale." https://www.wisdomtree.com/us/insights/blog/from-demo-to-distribution-what-meta-s-muse-tells-us-about-agents-at-scale
20. RuntimeWire, "Aaron Levie says agents will use software 100x more than people" (embeds x.com/levie/status/2068851573175021864). https://runtimewire.com/article/aaron-levie-says-agents-will-use-software-100x-more-than-people-and-force-new-sa
21. The New Stack, "Anthropic gives @Claude a permanent seat in your Slack channels," 23 June 2026. https://thenewstack.io/anthropic-claude-tag-slack/
22. Latent Space, "[AINews] Claude Tag: Multiplayer, Proactive, Persistent Agents in Slack" (AI Twitter recap citing @ClaudeDevs, @_catwu), 24 June 2026. https://latent.space/p/ainews-claude-tag-multiplayer-proactive
23. Dealroom / TBPN, "OpenAI DevDay 2026: Altman and Embiricos on dots." https://dealroom.co/news/talk-efcG5Uf-GH0-openai-devday-2026-altman-and-embiricos-on-dots-the-open-ecosystem-and-c/
24. AI Skill Market, "Teams of Dots: What Multi-Agent Coordination Looks Like in Practice." https://aiskill.market/blog/teams-of-dots-what-multi-agent-coordination-looks-like-in-practice
25. "When Can Human-AI Teams Outperform Individuals? Tight Bounds with Impossibility Guarantees," arXiv 2605.08710, May 2026. https://www.emergentmind.com/papers/2605.08710
26. AI Agent Store, "Human-AI Synergy Weekly AI News, 14 to 22 September 2026" (summarising arXiv study of 15 September). https://aiagentstore.ai/ai-agent-news/topic/human-ai-synergy/2026-08-25
27. Diginomica, "Together we are stronger: Capgemini on agentic AI and human teaming." https://diginomica.com/together-we-are-stronger-capgemini-agentic-ai-and-human-teaming
28. João Queirós, "How Sam Altman Uses Dots to Protect His Mornings and Build More" (summary of The Every Podcast, 30 September 2026). https://www.ai.joaoqueiros.com/blog/sam-altman-dots-mornings-prototypes-ultrafast