Fred Roadmap
Current releases
Swift 3.x is the production release. Weaver 5.x is feature complete and in validation. Kea 1.5 is obsolete.
Swift provides the runtime and SDK, document ingestion and retrieval, governed capabilities, evaluation and deployment tooling. Weaver consolidates these foundations around a common space model and clearer component responsibilities.
Architectural boundaries
Fred connects users, data and services within governed spaces. Domain services retain responsibility for business rules and records; applications provide the interfaces and process logic needed for a particular activity.
Approval of an agent's tool call and approval of a business milestone are separate operations, with distinct responsibilities and records.
Weaver priorities
Governed spaces
Consolidate ownership, access and resource lifecycle across organization, team, project and personal spaces. Documents, reusable instructions and assistant configurations need consistent rules for placement and use. A shared space does not make every conversation shared.
Prompts and skills
Build skill management on the Control Plane's space model, alongside prompts. Migrate the default skills out of fred-agents and align their packaging with the Agent Skills specification. Prompts express instructions; skills package procedures and supporting resources; capabilities expose executable operations. Skill instructions do not grant access rights. A skills marketplace can build on this foundation later.
Business applications
Provide practical capabilities for tasks such as working with complex spreadsheets and preparing presentations. Allow domain applications to combine these capabilities with their own interfaces and services, under the same access model. Users should be able to follow a business activity without managing its documents, instructions and progress across separate chats.
Reliable, resource-conscious execution
Use explicit workflows for predictable process rules and durable execution where work must survive interruptions. Introduce model reasoning where its contribution justifies its latency, cost and variability. Code execution is an optional capability subject to isolation, permissions and resource limits.
The same criteria apply to enterprise collaboration and smaller operational deployments: a clear interface, controlled resource use and traceable outcomes.
Under investigation: shared business progress
The working hypothesis is that collective work needs a small, explicit business state and a shared place to consolidate results. This state would record the current phase, references to accepted documents, decisions and transition history, independently of individual conversations.
For a bid team, several people and agents could prepare contributions in parallel. An authorized person would approve a gate against the consolidated deliverables; the application would record that decision and advance the process.
The design must address concurrent contributions, document revisions and approval authority. Where this state belongs, and which minimal contracts Fred should provide, remain open. This is not yet a completed Weaver feature.
Under investigation: from documents to governed knowledge
Knowledge is often equated with documents. Increasingly, it takes forms suited to AI: Markdown text, tables, wiki pages and results produced by agents. Fred already converts documents into such forms; the original file becomes one source among others.
Collections as the governed unit
The working hypothesis is that Fred governs collections owned by a space, such as a corpus, a wiki or a partition of a graph, together with their provenance. It does not govern each fact they contain: finer governance would be neither necessary nor efficient. Knowledge engines manage the content inside a collection.
Promotion: from a conversation to the organization
Knowledge produced with an agent starts private, in a conversation. It can then be promoted to a team, and from a team to the organization. Each promotion would copy an immutable snapshot, with its provenance, into a collection of the target space. A person with the appropriate role, or a policy defined for that space, decides the promotion.
An agent reuses only what has been promoted to the level where it works. This gives agents a governed way to collaborate through the data they produce, rather than only by exchanging messages. Promotion policies, and what happens to knowledge derived from content that is later withdrawn, remain open questions.
Knowledge engines built by others
Fred supplies a default knowledge stack for document search and tabular queries. Other teams, in the open-source community or in proprietary products, should be able to provide richer engines, for example graph-based ones with their own storage, without Fred becoming the provider of every storage technology.
The SDK would let such an engine work within a space under a precise role, without implementing multi-tenancy, roles or security itself. It writes within an admitted mandate, answers on behalf of the person who asks, and exposes its tools to agents as capabilities. Its storage remains a projection of Fred's catalogue: purge hooks and periodic reconciliation remove withdrawn content, and a rights check at query time ensures that a missed purge never becomes an access leak.
A Rust SDK, and the ability to run the embedded Rust agent outside the Control Plane, are anticipated for lighter deployments.
Developer integration
Maintain a documented path from a local extension to evaluation, deployment and operation. Agents, tools, document processors and application interfaces should integrate through explicit contracts, with examples that demonstrate access boundaries as well as successful execution.
Distributed execution
Keep direct agent calls distinct from durable task delegation. Select the execution mechanism according to duration, recovery and tracking requirements. Distribution should serve a concrete integration need.
Platform guidance
Separate operator-editable guidance from platform-owned instructions, and reduce redundant model context. Authorization and business invariants must be enforced by services; prompt wording alone cannot guarantee them.
Evaluation criteria
Assess each addition against a concrete business activity: task completion and output quality, clarity of shared progress, correct access enforcement, recovery after interruption, latency and resource consumption. Extensions should integrate without modifying Fred's core.
Fred remains an open-source foundation for business assistance, consuming and connecting domain services. A general-purpose ERP and a universal business workflow engine are outside its scope.
An invitation to contribute
We welcome contributions from researchers, students and industrial teams. Exploring use cases on an open-source foundation, with real data and systems already in production, offers a rich setting for connecting research questions with operational experience.
Bring a use case, develop a capability, share an evaluation or document what worked and what did not. These contributions help shape both Fred and our understanding of AI-assisted work. Join us on GitHub to start the conversation.