Fusion editing is a collaboration model in which people and AI agents read, write, comment, review, and hand work to one another inside the same live document.
Instead of asking an AI for text in a separate chat and pasting the answer into a document, the document becomes the shared operating surface. Human cursors, agent changes, comments, and handoffs remain visible where the work itself lives.
Dokki uses the term fusion editing for this model of real-time human–agent collaboration.
The short answer
Fusion editing is a collaboration model in which humans and AI agents work in the same live document or workspace. Agent changes remain visible and editable, permissions apply to both people and agents, and review happens inside the shared artifact instead of through copied chat output.
Fusion editing is real-time co-creation between humans and AI agents in a shared document, with visible contributions, shared context, permissions, and review.
A system qualifies as fusion editing when:
humans and agents act on the same canonical document;
changes appear without manual copy and paste;
agent work is attributable and reviewable;
permissions apply to both people and agents;
comments and handoffs stay attached to the work;
the document remains useful after the chat or agent run ends.
Why a new editing model is needed
Most AI writing workflows use a split interface:
the human keeps the real document in one application;
the AI receives a partial copy in a chat;
the model returns a draft;
the human pastes it into the document;
feedback happens in another message;
provenance disappears during the transfer.
This works for isolated generation. It breaks down when several people and agents contribute over time.
The context gap
An agent working in a separate chat sees only the context supplied for that session. It may miss the current document, comments, related tables, project decisions, or changes made by another teammate.
The handoff gap
The agent's output arrives as a block of text rather than a change inside the work. Someone must decide where it belongs, reconcile it with current content, and copy it over.
The accountability gap
Once AI-generated text is pasted into a document, reviewers may not know which agent produced it, what sources were used, or which instructions shaped the result.
The memory gap
The chat transcript preserves conversation, but the team needs a durable artifact: the approved brief, plan, specification, table, or publication.
Fusion editing closes these gaps by making the artifact—not the conversation—the shared center of work.
How fusion editing works
Fusion editing combines four layers.
1. A shared document model
Every participant operates on the same underlying document state. Real-time collaborative editors commonly use Conflict-free Replicated Data Types, or CRDTs, to merge simultaneous changes without forcing one user to overwrite another.
Dokki's collaborative editor uses Tiptap, Yjs, and Hocuspocus. Tiptap documents how Yjs provides a shared document model, while awareness data can expose presence, cursors, selections, and other real-time user states.
2. Agent access to document operations
The agent needs structured ways to find, read, and edit the document. It should not depend on a screenshot or a pasted excerpt.
Typical operations include:
search the workspace;
read a document or section;
insert content at a known location;
replace a section;
add or update table rows;
create a related document;
return the resource link for review.
In an open system, these operations can be exposed through the Model Context Protocol so external agents work with the same resources as the built-in agent.
3. Identity, permissions, and attribution
An agent should not become an invisible superuser. It needs an identity and a bounded scope.
A production system should answer:
Which human or agent made this change?
Which workspace and resource could it access?
Was the action read-only, write, delete, or publish?
Did the action require confirmation?
Can the credential be revoked?
Can a reviewer recover an earlier version?
4. Human review moments
Fusion editing does not mean autonomous acceptance. It makes review part of the shared artifact.
Useful review controls include:
comments on a specific passage;
suggested or scoped changes;
status fields such as Draft, Review, and Approved;
version history and restoration;
explicit confirmation before destructive or public actions;
named owners for unresolved decisions.
Fusion editing versus AI-assisted writing
AI-assisted writing is a feature. Fusion editing is an operating model.
Fusion editing versus human-only real-time collaboration
Google Docs-style collaboration lets people edit one document together. Fusion editing adds software agents as first-class participants.
That requires more than displaying another cursor:
agents must understand task and workspace boundaries;
their edits need structured APIs or tools;
credentials must be scoped and revocable;
long-running work needs status and handoff behavior;
the system must distinguish retrieved content from trusted instructions;
public or destructive actions need additional control.
The real-time document layer solves simultaneous editing. The agent-governance layer makes machine participation safe and useful.
Fusion editing versus autonomous agents
An autonomous agent may work alone and deliver a final result. Fusion editing favors visible collaboration around a shared artifact.
Both patterns are useful. Use autonomous execution for bounded background work. Use fusion editing when judgment, iteration, shared ownership, or auditability matters.
The core collaboration loop
A practical fusion-editing loop has six stages.
1. Frame
A human defines the outcome, constraints, source rules, and review owner in the document.
2. Research
An agent searches approved workspace and external sources, then attaches evidence to claims or a structured table.
3. Draft
The agent adds content directly to the target document or creates a clearly linked draft.
4. Review
A human comments, edits, rejects weak claims, or changes the direction without losing the surrounding artifact.
5. Revise
The same or another agent resolves comments, rechecks sources, and updates only the affected sections.
6. Approve and reuse
The team marks the artifact approved, publishes or hands it off, and keeps it available as context for future work.
The loop can repeat at section level rather than regenerating the entire document.
A concrete example: product launch brief
Imagine a startup preparing a product launch.
Human contribution
The product lead creates the brief, states the target audience, locks the launch date, and marks pricing as requiring executive approval.
Research-agent contribution
An agent searches customer interviews, competitor updates, and prior launch retrospectives. It builds an evidence table and adds source-backed findings to the brief.
Writing-agent contribution
A second agent drafts positioning, FAQs, release notes, and channel variants using the approved evidence.
Human review
Marketing edits the narrative. Legal comments on claims. Product corrects feature scope. Each review remains attached to the relevant content.
Release-agent contribution
After approval, an agent prepares publishable assets and updates the launch tracker. Publication still requires explicit confirmation.
The shared brief becomes the control plane for research, writing, review, and release.
What good fusion editing feels like
The document is always current
People do not wonder which chat contains the latest answer. The canonical artifact reflects accepted work and visible drafts.
Agent activity is legible
Reviewers can identify what the agent changed, which task it was executing, and where the evidence came from.
Work can change owners
A human starts a section, an agent expands it, another human challenges it, and a second agent verifies it. The artifact survives every handoff.
Context stays bounded
An agent receives the workspace and resource scope required for the job, not indiscriminate access to the organization.
Review is proportional to risk
Low-risk drafting may proceed directly. Deletion, external messages, credential changes, and public publishing require confirmation.
Design principles for fusion-editing systems
One canonical artifact
Do not create a new disconnected copy for every agent run. Make the target explicit and return its resource link.
Visible actor identity
Agent edits should carry a recognizable identity, not impersonate the human operator.
Structured, minimal edits
Prefer section-level insertion or replacement over rewriting an entire document. Smaller changes are easier to review, merge, undo, and attribute.
Workspace-scoped access
Permissions should be granted at the smallest useful boundary: workspace, folder, document, or operation.
Source-linked claims
Research contributions should preserve the source URL, document reference, or evidence row. Fluent prose is not provenance.
Recoverability
Version history, snapshots, or other restoration mechanisms are necessary because people and agents can both make mistakes.
Explicit state
Use statuses, owners, and review gates. Do not infer that a polished paragraph is approved.
Durable handoffs
A handoff should state what changed, what remains open, who owns the next step, and where the work lives.
The technology behind real-time fusion editing
CRDT document state
A CRDT-based editor represents changes in a way that can be merged across participants. Tiptap's collaboration stack uses Yjs and documents support for simultaneous edits, offline work, and synchronization.
Awareness and presence
Presence is separate from durable document content. Awareness data can show active participants, cursor positions, selections, or custom states without writing every movement into document history.
Collaboration server
A provider or collaboration server distributes updates, authenticates connections, and persists the shared state. Dokki uses Hocuspocus as its real-time collaboration service.
Agent tool layer
Agents need semantic operations above raw CRDT updates: read this section, insert after this node, update these cells, search this workspace, or publish this resource.
Knowledge and retrieval layer
Workspace search lets an agent find related documents and tables without loading the entire organization into context. Retrieval results still require source and permission checks.
Audit and version layer
Snapshots and activity metadata allow teams to inspect and restore work. Real-time merge correctness does not replace governance.
Permissions and safety
Fusion editing increases the number of capable participants. Its safety model should be explicit.
Read is not write
An agent that summarizes a workspace may not need document-edit privileges.
Write is not publish
Drafting inside a private workspace is lower risk than making content public or pushing it to an external domain.
Content is not instruction
Retrieved documents, websites, and comments may contain prompt injection. Agents must treat them as data unless a trusted policy says otherwise.
Identity is not authority
An agent may be installed by an administrator but still operate with narrow task permissions.
Confirmation is not friction everywhere
Require confirmation where impact is material: deletion, broad moves, public sharing, external messages, payments, credentials, or irreversible changes.
Metrics for human–agent collaboration
Measure the workflow, not only model output.
A high generation rate with low acceptance or heavy rework is not successful collaboration.
When to use fusion editing
Fusion editing works well for:
research briefs and competitive analysis;
product requirements and technical specifications;
launch plans and editorial workflows;
policies and operating procedures;
proposals and client deliverables;
incident reviews and decision logs;
living documentation that agents maintain over time.
When not to use fusion editing
Use a simpler workflow when:
the task is a disposable personal question;
the output has no value after the conversation;
only one short transformation is needed;
the agent cannot be given safe access to the artifact;
regulation requires a separate controlled authoring environment;
the work is better represented as code with established version-control review.
Fusion editing should reduce handoff cost, not add ceremony to trivial tasks.
A rollout checklist for teams
Phase 1: read and comment
connect one low-risk workspace;
allow agents to search and read;
have agents return source-linked comments;
measure whether reviewers can verify the output.
Phase 2: scoped drafting
allow creation in a sandbox folder;
define draft and review statuses;
require resource links in every handoff;
test version restoration.
Phase 3: controlled edits
allow section or table-row updates;
preserve agent identity and activity;
add confirmation for destructive operations;
monitor accepted edit rate and rework.
Phase 4: recurring collaboration
schedule maintenance and monitoring tasks;
assign connector owners and rotation dates;
review stale sources and permissions;
separate draft, approval, and publishing authority.
Frequently asked questions
What is fusion editing?
Fusion editing is real-time collaboration in which humans and AI agents work inside the same live document, with shared context, visible contributions, permissions, comments, and review.
Is fusion editing the same as collaborative editing?
It extends collaborative editing. Traditional collaborative editing connects multiple people. Fusion editing also treats AI agents as participants that can read and modify the shared artifact through governed tools.
Is fusion editing just AI writing?
No. AI writing generates or transforms text. Fusion editing coordinates ongoing work between people and agents in a canonical, reviewable artifact.
Can multiple agents work on the same document?
Yes, if the system handles concurrent state, gives each agent a scoped identity, and prevents overlapping tasks from causing uncontrolled rewrites. Section ownership and structured edits reduce conflict.
How do you review an AI agent's edits?
Keep edits inside the document, preserve actor and source information, use comments and statuses, compare versions where needed, and require explicit approval for high-impact actions.
What makes fusion editing safe?
Workspace-scoped permissions, distinct agent identities, minimal edit operations, source boundaries, prompt-injection defenses, confirmation gates, activity history, and version restoration.
Does fusion editing replace chat?
No. Chat remains useful for coordination, clarification, and lightweight requests. Fusion editing ensures that durable work and decisions live in the artifact rather than only in the conversation.
What is an example of human–agent collaboration?
A product lead frames a launch brief, a research agent adds sourced evidence, a writing agent drafts messaging, teammates comment in the document, and a release agent prepares approved content for publication.
Final definition
Fusion editing is the shift from AI beside the document to AI inside the collaborative document workflow.
The defining feature is not automatic text generation. It is shared ownership of a live artifact: people and agents can see the same work, contribute within their permissions, hand it off without losing context, and leave behind a result that remains reviewable and reusable.