From Session Fragments to Shared Truth

This beta release changes something fundamental in Masterminds: your work no longer has to stay trapped inside isolated sessions. Shared initiative context is now real, agents can run longer inside a single processing window, and long conversations are much easier to sustain without sacrificing control.

The Old Bottleneck

Before this release, a lot of valuable work still behaved like session-only memory. You could explore multiple paths, test competing theses, refine personas, and build alternative roadmaps, but once one direction clearly won, promoting that work across the initiative was still too manual.

That created friction at exactly the wrong moment.

The more seriously you used Masterminds, the more likely you were to hit it:

• strong work was spread across multiple sessions;
• you had to re-upload or re-import material to make it reusable;
• long-running agent flows became harder to sustain as conversations accumulated more documents;
• collaborative work suffered when context, tools, and stop controls were not durable enough.

This release addresses that whole cluster, not with one isolated fix, but with a better operating model for initiative-level work.

Shared Initiative Context Is the Big Shift

The most important change is simple to describe and powerful in practice: when a document becomes initiative-level context, every session inside that initiative can use it.

That matters because real product work is not linear. You usually explore multiple routes before you find the winning one. A thesis wins. A persona wins. A roadmap wins. A research artifact becomes the one source of truth.

Now, instead of recreating that artifact in every session, you can promote it into the initiative scope and make it available across the initiative.

In practice, that promotion is direct: drag the document to Initiative in the right-side panel to move it there. Hold Option on Mac or Ctrl on Windows while dragging if you want to copy it instead.

Precedence Still Matters

Shared context is not a blunt overwrite. It follows a clear hierarchy:

1. Session
2. Initiative
3. System

If a document with the same name exists in both session and initiative, the session version wins.

This is not a bug. It is what allows safe experimentation.

You can promote a winning artifact to the initiative so everyone can use it, while still keeping a local session override when you want to test a variation without changing the shared baseline. The trade-off is that naming now matters more. If you promote a document and an identically named session document already exists, the session document will continue to be the one the runtime uses in that session.

That is the real mental model:

• Initiative scope = shared truth for the initiative
• Session scope = local override for active experimentation

Longer Conversations Need Better Controls

Shared context only helps if long-running sessions remain usable. That is the second big theme of this release.

Masterminds now gives you explicit control over how much recent history gets sent into the next turn. Next to the model selector, you can choose how much history to include:

• All (default)
• 20
• 10
• 5

This is not just a token-saving toggle. It changes how large initiatives remain workable over time.

When a session accumulates a lot of documents and a long message history, you can now reduce the number of recent turns that travel into the next request, while watching the context gauge reflect the reduction. That makes the trade-off visible: less recent message history, less context consumption, longer usable conversations.

The result is practical, not theoretical:

• sessions with many documents keep moving;
• you can deliberately extend conversation length instead of starting over too early;
• the platform makes context pressure visible instead of hiding it.

Agents Now Run Longer Inside a Single Processing Window

This release also strengthens the runtime itself.

Agents can now sustain longer autonomous processing, including multiple tool calls within a single run. Sequential execution is stronger, long flows are less likely to skip steps, and agentic loops have been refined to behave more consistently under heavy context.

That is why this release is not just about “shared documents.” It is also about shared execution confidence.

The AI response path is more stable, double-processing issues were addressed, agentic loop behavior was tightened, and the system is currently running at roughly 50% faster performance on this beta path, with more optimization work already underway.

For users, the visible outcome is straightforward:

• fewer broken chains in multi-step execution;
• fewer cases where the AI appears to process twice;
• fewer moments where long flows feel fragile;
• better continuity when the work requires several actions, not one answer.

Vibe Mode Becomes a Real Exploration Tool

Vibe Mode is now much more meaningful for fast exploration.

In practical terms, it can now support autonomous execution across roughly five to ten steps, depending on the flow and the quality of the context you provide. That makes it useful for quickly testing alternatives, generating first passes, and exploring structured paths without hand-holding every step.

It still comes with a caveat, and that caveat matters.

When the system is allowed to run more autonomously, some generated documents may still land below the quality bar you would expect from a carefully reviewed, manually directed execution. Vibe Mode is not a replacement for judgment. It is a speed tool.

Used correctly, that is exactly its value:

• fast exploration of alternatives;
• faster first drafts;
• less friction in process-heavy flows;
• more momentum before you decide where to slow down and review deeply.

Practical Enablers: Search, Uploads, Downloads, and MCP Sharing

The release is stronger because the surrounding tools also got better.

Search That Helps You Recover Lost Work

You can now search sessions more effectively and search inside variable content, which makes it much easier to recover the “where did that go?” moments that happen in real projects.

That sounds small until your initiative has dozens of sessions and a deep variable tree.

Uploads That Respect Workflow

Uploads are now reliable across the three workflows people actually use:

• during session creation,
• from the paperclip in chat,
• from the right panel upload control.

The right-panel upload path is especially important because it lets you add documents to context without triggering AI immediately. You can preload the session, organize the material, and let the agent use it naturally on the next turn.

Downloads That Match the Mental Model

Downloads are now much more consistent across:

• single document vs. full tree;
• session scope vs. initiative scope vs. both;
• multiple export formats.

That is the difference between “the content exists in the platform” and “the content is safely portable.”

MCP Sharing at the Initiative Level

MCPs can now be promoted to the initiative so other members of that initiative can see and use them.

There is still an important limitation today: compatibility is strongest with bearer-token setups, so some MCPs that depend on richer auth models still need follow-up work. That fix is already on the roadmap. But the sharing model itself is now much closer to how teams actually operate.

Collaboration and Editing Feel More Durable

Some of the most valuable improvements are the ones you feel without naming them.

• collaborator invites have been restored through Team Collaboration;
• The Masterminds Operator Inline is much more reliable for quick edits and focused questions on a document;
• AI-driven updates to existing documents are now versioned automatically;
• the agent selector is easier to scan, with clearer grouping across Masters, Minds, Operators, and custom agents.

Two caveats remain worth stating plainly:

1. The Masterminds Operator Inline still has a known issue in Diff Mode acceptance, where some updates are not yet applied correctly. Immediate accept flows are much stronger.
2. Leaderboard / StageFlow / Kanban are in a better state, but alignment with agent-flow logic is still evolving.

That is the right way to read this beta: major progress, with a few clearly identified follow-up improvements already underway.

Why This Release Matters

Most AI products still behave like glorified conversation logs. Masterminds is moving in a different direction.

The platform is becoming an initiative system, not just a session system.

That means:

• knowledge can graduate from local exploration to shared operational truth;
• long-running work stays viable for longer;
• autonomous execution becomes more useful before manual intervention;
• collaboration gets less brittle;
• context is easier to load, search, reuse, export, and trust.

That is the real upgrade.

Not just “more features,” but a better continuity model for serious product work.

What To Do Next

If you are already using Masterminds, the best way to feel this release is:

1. promote a winning document to initiative scope;
2. open another session in the same initiative and verify how shared truth works;
3. reduce history from All to 20, 10, or 5 and watch the context gauge change;
4. try a longer multi-step run with Vibe Mode on a non-critical flow;
5. preload documents through the right panel before asking the agent to work;
6. use search to recover a session or variable you have not touched in days.

And if something breaks, feels slow, or still feels harder than it should, use the Feedback button in the top-right area of the app. The team is reading submissions, tracking votes, and using that signal to prioritize the next wave of improvements.

Also: the Academy now has a growing body of documentation covering the platform, agent order, workflows, and operating concepts. If you want the mental model behind the product, not just the UI, start there.

What’s Next

• better MCP auth compatibility beyond bearer-token-first setups;
• more refinement in StageFlow and Kanban alignment with agent flows;
• continued speed work on the AI runtime;
• ongoing refinement of existing agents and new agents now in progress.

This beta release closes a real gap. The next one should make the system feel even less fragmented, more autonomous, and easier to trust at initiative scale.