Hub Mode is the name Anthropic has attached to an unreleased control surface for Claude’s sub-agents, and it surfaced the way these things usually do: not in a blog post, but in a settings panel that was never meant to be seen. On 18 August 2026 the feature-tracking site TestingCatalog reported a new option in the Claude mobile app, sitting next to a planet icon and flagged as internal. Anthropic has said nothing about it.

The description attached to the toggle is short and slightly cryptic. It “will apply to new tasks and requires users to restart the app after activation.” Turning it on produced no visible change at all, which is the signature of a feature whose interface has shipped ahead of its backend. That pattern is common at Anthropic — the same staging behaviour preceded enterprise-managed auth for Claude MCP connectors, which spent weeks as inert scaffolding before it became generally available.

What makes this one worth a closer look is the second sighting. Claude Code on the web has a Hub destination of its own, and opening it produces a list of sub-agents. Select one and the page states that Claude will generate a task board for the user and their agents. A task board is a very specific promise. It implies persistent work items, assignment, and status — the parts of agent orchestration that Anthropic has so far kept inside a terminal, a script variable, or a JSON file on your own disk.

This breakdown covers exactly what has been observed and where, why the planet icon has a history, what Hub Mode would sit on top of in Anthropic’s existing orchestration stack, how it compares with the rival agent consoles shipping now, what remains genuinely unknown, and what teams running agents at scale should do before it arrives. Every figure below is traceable to a document in the References section.

Hub Mode: What Anthropic Is Actually Testing

hub mode anthropic claude task orchestrator b ladder five rungs

The honest summary is that Hub Mode is a name, an icon, a toggle and one sentence of UI copy. That is less than a product and more than a rumour, and the distinction matters when you are deciding whether to plan around it.

The mobile sighting

The toggle lives in the Claude mobile app’s settings, marked internal — Anthropic’s convention for options exposed to its own staff and visible in shipped builds only by accident. It is accompanied by a planet icon. The copy promises the mode “will apply to new tasks,” which is the first hint that Hub Mode is scoped to task execution rather than to chat.

The Claude Code sighting

Separately, Claude Code on the web carries a Hub destination that lists sub-agents. Choosing one shows a message about generating a task board for the user and their agents. Neither the mobile toggle nor the web destination currently does anything observable.

What is not confirmed

There is no confirmed link between the two. They could be one system with two entry points, or two unrelated features that happen to share a name. Anthropic has published no documentation, no changelog entry and no release date for Hub Mode, and the company did not comment when the sighting was reported.

ObservationSurfaceEvidenceStatus
Hub Mode toggleClaude mobile app settingsOption marked internal, planet iconVisible, no effect when enabled
“Applies to new tasks”Toggle descriptionRequires an app restart after activationCopy only
Hub destinationClaude Code on the webLists sub-agents; offers a task boardNavigable, non-functional
Planet icon lineageMobile app, earlier buildsSame icon previously labelled OrbitOrbit label removed
Official positionAnthropicNo blog post, docs or changelogUnannounced

Where Hub Mode Was Found, and Why the Wording Matters

hub mode anthropic claude task orchestrator c solid anvil

Feature archaeology is guesswork, but the wording narrows the guesswork considerably. Three phrases in the Hub Mode evidence do real work.

“Will apply to new tasks”

Tasks is not a casual noun at Anthropic. It is the name of a feature already being prepared for the mobile apps: an iOS build was spotted with a Tasks entry in the app menu and a dedicated Tasks page for creating repeatable jobs, mirroring Claude Cowork’s naming and iconography. Strings in the same build hinted that Claude could operate a browser as part of executing one. Hub Mode applying “to new tasks” reads as Hub Mode changing how those jobs are run.

“A list of sub-agents”

Sub-agents are the worker primitive across Claude Code. A Hub that lists them is a directory, not a chat. Directories imply identity, reuse and state that survives a single conversation — none of which the current sub-agent model gives you on mobile.

“Generate a task board”

Boards are the piece Anthropic has never shipped to a general audience. Claude Code has a shared task list for agent teams, but it lives in ~/.claude/tasks/ on the machine that ran the session, is never uploaded, and is scoped to exactly one team per session. A generated task board on the web would be the first time that structure left the local filesystem.

From Orbit to Hub Mode: The Planet Icon Has Previous

hub mode anthropic claude task orchestrator d rack three shelves

The planet icon attached to Hub Mode is not new. The same icon previously appeared under the name Orbit before that label was removed, which is the strongest available hint that Hub Mode is an evolution of an earlier idea rather than a fresh one.

What Orbit was

Orbit was reported in May 2026 as a proactive briefing tool for Claude Cowork: opt-in, timezone-aware, generating personalised briefings from connected work applications. The initial integration list was Gmail, Slack, GitHub, Calendar, Drive and Figma. Its own settings placeholder read “Your deployed Orbit apps. Pin favorites for quick access.”

Why the rename is interesting

Orbit was framed around information arriving on its own. Hub Mode is framed around agents being coordinated. If the icon carried over, the underlying surface may have been repurposed from a briefing feed into an orchestration console — a meaningful change of ambition, and one that fits the direction of travel across the whole industry, where orchestration has become the central customer experience problem rather than a back-office detail.

How long these things sit in the oven

Orbit was reported on 4 May 2026. Hub Mode was reported on 18 August 2026 — 106 days later. Neither has shipped. The chart below sets that gap against the 10 days between the Hub Mode sighting and this article.

Internal-build sightings: elapsed days (source dates 4 May, 18 Aug, 28 Aug 2026)
Orbit sighting to Hub Mode sighting — 106 days
Hub Mode sighting to this article — 10 days
Bars are each interval as a share of the longer 106-day gap.

What Hub Mode Would Sit On Top Of

hub mode anthropic claude task orchestrator e dumbbell bar two discs

Anthropic already has four distinct ways to run more than one agent. Hub Mode is interesting precisely because none of them currently has a shared visual home, and a hub is what you build when you have four things that need one.

Sub-agents

The lightweight primitive. Claude spawns a worker, the worker does a focused job, the result returns to the caller’s context. Coordination is entirely Claude’s, turn by turn, and intermediate results land in a context window.

Agent teams

Experimental and off by default — you enable them with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. One session leads, teammates run as separate Claude Code instances with their own context windows, and they coordinate through a shared task list plus a mailbox of JSON files. Anthropic’s own guidance is to start with three to five teammates and around five to six tasks each.

Dynamic workflows

The heavyweight option: a JavaScript script Claude writes, executed by a runtime in the background while your session stays responsive. Available on all paid plans from Claude Code v2.1.154, and on Bedrock, Google Cloud’s Agent Platform and Microsoft Foundry.

Managed agents on the platform

On the Claude Developer Platform, a coordinator delegates to agents that share one sandbox, filesystem and vault credentials, but each runs in its own context-isolated session thread. Threads persist, so the coordinator can send a follow-up to an agent it called earlier.

PrimitiveWho holds the planWhere results liveScale per runAvailability
Sub-agentsClaude, turn by turnClaude’s context windowA few per turnGeneral
Agent teamsThe lead agent, turn by turnA shared task list on diskA handful of peersExperimental, off by default
Dynamic workflowsThe scriptScript variablesDozens to hundredsPaid plans, v2.1.154+
Managed agentsThe coordinatorPersistent session threadsCoordinator plus specialistsBeta on the platform
Hub ModeUnknownA generated task boardUnknownInternal only

Hub Mode and the Numbers Behind Agent Orchestration

hub mode anthropic claude task orchestrator f archway two posts crossbar

Anthropic publishes hard limits for the orchestration it has already shipped, and those limits are the best available guide to the scale a control surface would need to display.

The caps inside one workflow run

A dynamic workflow runs up to 16 concurrent agents, fewer when Claude Code has fewer CPUs available. The lifetime ceiling is 1,000 agents per run, a runaway-loop backstop set far above any real workflow, and a single fan-out call accepts at most 4,096 items.

The warnings that fire before you hit them

Claude Code flags a run that schedules more than 25 agents, or whose projected token total passes 1.5 million, with a Large workflow warning in the task panel. The warning is advisory: it does not pause or limit the run.

The size guideline that shapes what Claude writes

A separate setting tells Claude how many agents to aim for: small is fewer than 5, medium — the default — is fewer than 15, and large is fewer than 50. It is advice sent to the model, not a cap.

Dynamic workflow size guideline: agents Claude aims for
small — fewer than 5
medium (default) — fewer than 15
large — fewer than 50
Bars are each guideline as a share of the largest, 50 agents.
Concurrency versus the lifetime ceiling in one workflow run
Running at any one moment — 16 agents
Advisory warning threshold — 25 agents
Hard cap for the whole run — 1,000 agents
Bars are each figure as a share of the 1,000-agent lifetime cap. 16 concurrent agents is 1.6% of it.

Why a Task Board Is the Missing Piece

Every orchestration primitive above answers the question “how do I run more agents.” None of them answers “where do I look while they run, and what is left when they finish.” That gap is what a Hub Mode task board would close.

Results that vanish with the context

Sub-agent results land in a context window and are gone when the conversation ends. A workflow’s intermediate results live in script variables the runtime discards. Neither leaves an artefact a colleague can open tomorrow.

State that never leaves one machine

The agent-teams task list is the closest thing to a board Anthropic ships, and its limitations are documented plainly: the directory persists locally and is never uploaded, there is one team per session, teams cannot nest, and /resume does not restore in-process teammates. Task status can also lag, because teammates sometimes fail to mark work complete and block whatever depended on it.

Mobile has nothing at all

On a phone there is no terminal, no /workflows progress view and no agent panel. If Anthropic wants agent work initiated or supervised from mobile — and a Tasks page in an iOS build suggests it does — a board is the only sensible interface. Hub Mode arriving on mobile first is therefore a clue about intent, not an accident of where it was spotted.

What a board needs to showExisting plumbingGap Hub Mode would close
Which agents existSub-agent definitions in .claude/agents/No cross-device directory
What each is working onShared task list, pending / in progress / doneLocal only, one team per session
What blocks whatTask dependencies resolved automaticallyInvisible outside the terminal
Live progress and cost/workflows view: phases, tokens, elapsed timeDesktop and CLI only
What survives the sessionTask directory under ~/.claude/tasks/Never uploaded, never shared

How Hub Mode Compares With Rival Agent Consoles

Anthropic is not first to the idea of a control room for agents, and the comparison sets a floor for what Hub Mode has to do to matter.

The competitive shape

Every major assistant vendor is converging on the same three ingredients: a roster of named agents, a queue of durable jobs, and a place to watch them. The differentiator in artificial intelligence tooling is no longer whether an assistant can spawn workers — it is whether a person can supervise them without reading a transcript.

Where Anthropic starts from strength

Between Claude Code, Cowork, managed agents and the Agent SDK, Anthropic has arguably built more agentic tooling than any rival. Hub Mode’s opportunity is not new capability. It is coherence — one place where the four primitives above stop looking like four products.

Where the risk sits

The risk is the same one every dashboard faces: a board that shows work without letting you change it is a status page, not a control centre. The one observed string says Claude “will generate” the board, which suggests it is derived from agent activity rather than authored by the user. Whether Hub Mode lets you reassign, re-prioritise or stop work is completely unknown.

The cybersecurity question a remote board raises

Any surface that can start or steer agents from a phone inherits a permissions problem, and cybersecurity is the part of agent orchestration Anthropic has been most careful about so far. In agent teams, teammate permission prompts deliberately bubble up to the lead session for a human to approve.

A teammate cannot approve a prompt or supply consent on your behalf, and one that was denied an action cannot relay it to another teammate to get around the check. In auto mode the classifier goes further, treating an approval claim relayed from another agent as untrusted input and reviewing every inter-agent message before delivery. A Hub Mode board would have to preserve all of that, on a device with far less context about what is being approved.

What Remains Unknown About Hub Mode

It is worth being explicit about the size of the unknown, because coverage of internal builds tends to round speculation up into fact.

No date, no tier, no scope

Anthropic has published no release timeline, no indication of which plans would get Hub Mode, and no statement of whether it targets individuals, teams or enterprise administrators. The mobile toggle’s internal marking means even the staged-rollout population is unclear.

No confirmed relationship between the two surfaces

The mobile toggle and the Claude Code web destination share a name and nothing else that has been verified. Treating them as one feature is an assumption, and it may prove wrong.

No evidence it works yet

This is the most important caveat. Enabling Hub Mode changed nothing observable, and the web destination produced no functioning board. Features at this stage are routinely cut. The Orbit label disappearing from the same icon is a direct precedent within the same app.

What to Do Before Hub Mode Ships

None of this is a reason to wait. The work that makes a future control surface useful is the work that makes your current agent runs legible, and it pays off either way.

Write your agents down

If Hub Mode lists sub-agents, the list will be built from definitions that already exist. Teams whose agents live in reusable definitions — with a name, a model, a tool allowlist and a written brief — will get a populated directory. Teams improvising prompts per session will get an empty one.

Make the work items real

Boards need discrete, checkable items. Anthropic’s own advice for agent teams is a good proxy: self-contained units producing a clear deliverable, roughly five to six per teammate, sized so coordination overhead does not exceed the benefit. That is a discipline worth adopting now, at the point where you brief the work. Our autonomous AI agents practice covers the same ground for business processes rather than codebases, and the same task-shaping habits carry straight over from intelligent automation projects that never involved a model at all.

Decide what a human still signs off

A board makes delegation easier, which makes it easier to delegate too much. The teams that get the most out of autonomous agents are the ones that decided in advance which steps a person approves — the same discipline that keeps workflow automation honest. Write that list before a control surface arrives, not after.

Instrument cost before it is someone else’s problem

Multi-agent runs cost meaningfully more than a single conversation, which is why Claude Code warns at 25 scheduled agents and 1.5 million projected tokens. Whoever ends up watching a Hub Mode board will be watching spend as much as progress. Knowing your current per-run baseline is what makes that number mean anything.

Do not build against an unreleased surface

Plan for the primitives that exist and are documented — sub-agents, agent teams, dynamic workflows, managed agents. Hub Mode is scaffolding in a shipped build, not an API. Building around it today is building around a screenshot.

Hub Mode: Frequently Asked Questions

Is Hub Mode available to the public?

No. It appears as an internal option in the Claude mobile app and as a non-functional destination in Claude Code on the web. Enabling it produces no observable change.

Has Anthropic confirmed Hub Mode exists?

Anthropic has made no announcement. The evidence is entirely from strings and interface elements found in shipped builds and reported by TestingCatalog on 18 August 2026.

Is Hub Mode the same thing as Orbit?

Unconfirmed. The two share a planet icon, and the Orbit label was removed from that icon before Hub Mode appeared, which suggests a lineage rather than proving one. Orbit was described as a proactive briefing tool; Hub Mode is described in terms of tasks and sub-agents.

Will Hub Mode replace dynamic workflows or agent teams?

Nothing suggests replacement. A hub is a place to see and steer work, not a new way to execute it. The most likely reading is a control surface over primitives that already exist.

What can I use today instead?

Dynamic workflows are generally available on paid plans from Claude Code v2.1.154, with /workflows giving you phases, agent counts, token totals and elapsed time. Agent teams give you a shared task list once you set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1, with the limitations noted above.

References and Further Reading