November 8, 2026 in platform, architecture by Dimitri Tombroff6 minutes
An analysis of governed spaces, reusable expertise and the separation between agent execution and business processes in Fred Weaver and four major platforms.
Update, November 18, 2026: Weaver is now feature complete and in validation. See the Weaver release page for its current status.
Fred is an open-source environment for combining data, tools and AI assistance within governed business activities. Weaver is its next major release, currently under development and available as an incomplete developer preview. It builds on Fred’s existing foundations to support shared workspaces, reusable resources and a clear separation between access control, document processing and agent execution.
This article examines that architecture alongside selected mechanisms from OpenAI, Anthropic, Microsoft and Salesforce. The comparison concerns architectural responsibilities, rather than feature coverage or measured performance. It identifies convergent design choices and the questions Fred is investigating.
In Fred, organizations, teams and projects define where resources belong and who may use them. A common space model also covers personal resources. Access to a collective space does not expose another person’s conversations, and roles do not automatically propagate to child spaces.
flowchart TB
subgraph organization["Organization"]
direction TB
common["Common knowledge
Assistants and resources"]
subgraph team["Team"]
direction TB
shared["Team knowledge
Assistants and resources"]
subgraph project["Project"]
focused["Resources for
a shared objective"]
end
end
common -->|"According to access rights"| shared
shared -->|"According to access rights"| focused
end
classDef neutral fill:#F1F5F9,stroke:#64748B,color:#0F172A
class common,shared,focused,organization,team,project neutral
Figure 1. Collective spaces and access to resources. Reproduced from the Weaver architecture introduction.
This model gives Fred a starting point for collective work that is independent of any particular agent or conversation.
The Agents SDK provides tool execution, specialist handoffs, conversation continuity and resumable approval interruptions. Specialists can take over execution through handoffs or contribute as tools while a coordinating agent retains control.
Fred addresses these execution concerns within a wider scope: several people and agent runs may contribute to the same business activity. Persisting an agent run provides execution continuity; the application’s shared progress and decisions require their own representation.
Agent Skills packages instructions, scripts and supporting resources that an assistant can discover and load as needed. This separates reusable expertise from the implementation of a particular agent. Anthropic’s guidance on effective agents also distinguishes predefined workflows from model-directed execution.
This approach is consistent with Fred’s use of reusable resources and capabilities. Weaver’s additional concern is their placement and governance: which space owns a resource, who may maintain it, and which tools the executing agent is authorized to use. A skill’s instructions do not themselves confer access rights.
Copilot Studio supports rule-based flows alongside generative orchestration. Agents can select tools and invoke flows; flows can include agent calls and human approvals. A predictable control structure can therefore contain tasks whose outputs remain probabilistic.
Fred adopts the same separation of responsibilities as a design principle for business applications: process rules belong to the application, while agents assist with selected tasks. Those applications may integrate existing customer services or supply domain-specific extensions.
Agentforce connects agents to actions implemented through Flow, Apex and connected services. Asynchronous integration patterns allow downstream services to own execution independently of an agent session. Access depends on the running identity and configured permissions; custom actions must preserve the relevant checks.
Fred similarly exposes service operations through capabilities. Its architectural objective is to connect those services within governed spaces while retaining their ownership of business transactions and records.
When someone asks an assistant to work with documents, Fred must check what that person may access, run the assistant with the appropriate tools, and retrieve relevant content. Weaver separates these responsibilities into three services:
The diagram shows how these services cooperate during an interaction with an assistant.
flowchart TB
ui["User interface
Work in a selected space"]
cp["Access and resources
(Control Plane)"]
pod["Assistant execution and tools
(Agent Pod)"]
kf["Document processing and search
(Knowledge Flow)"]
ui -->|"Manage resources"| cp
ui <-->|"Chat"| pod
pod -->|"Check access and obtain context"| cp
pod -->|"Search authorized documents"| kf
cp -.->|"Request document processing"| kf
classDef fred fill:#DBEAFE,stroke:#2563EB,color:#172554
class ui,cp,pod,kf fred
Figure 2. Access control, assistant execution and document processing. Simplified from the Weaver system overview.
These component boundaries support a corresponding distinction in business applications:
| Concept | Responsibility |
|---|---|
| Prompt | Express instructions. |
| Skill | Package a reusable procedure and supporting resources. |
| Capability | Expose operations through tools. |
| Agent | Use instructions and tools to perform a task. |
| Business application | Maintain process rules, shared progress and decisions. |
For a bid team, agents may review requirements or prepare a cost analysis while the application records approved gates and accepted deliverables. Approval of a tool call concerns an execution; approval of a gate concerns the collective business process.
A durable shared dossier is a Weaver direction under investigation. It would retain current progress, references to contributions and human decisions independently of individual conversations. This extends Fred’s existing space model; it is not presented here as a completed workflow feature.
The comparison places Fred within a broader convergence toward reusable expertise, explicit processes and governed execution. Fred’s contribution lies in their composition within an open-source working environment. The remaining design question is how little additional machinery is needed to support collective business progress reliably. Weaver’s release compass records that direction.
The summaries below are paraphrases, not verbatim quotations. Documentation consulted on November 9, 2026.
OpenAI (Agents SDK & Human-in-the-Loop): The SDK provides an execution loop with tools and handoffs, conversation continuity, and approval interruptions that can be resumed from saved state. See Running agents, Orchestration and handoffs, and Guardrails and human review.
Anthropic (Agent Skills & Architecture): Skills package instructions, scripts and supporting resources that agents can load as needed. Anthropic also recommends starting with simple solutions and adding agentic complexity only when needed. See Equipping agents for the real world with Agent Skills (October 16, 2025; open-standard update December 18, 2025), Building effective agents (December 19, 2024), and the Agent Skills specification.
Microsoft (Copilot Studio & Hybrid Flows): Copilot Studio combines rule-based flows with generative orchestration that selects tools, topics, knowledge and agents dynamically. Flows can include agent calls and human approvals. See Agent flows overview and Orchestrate agent behavior with generative AI. Keeping critical business rules and authoritative record updates in explicit services or workflows is our architectural takeaway from these mechanisms.
Salesforce (Agentforce & Platform Architecture): Agentforce integrates agents with enterprise data and business actions, including Flow and asynchronous processing patterns. Access depends on the execution identity and configured permissions; custom actions must preserve the relevant security checks. See Agentic Integration Patterns and Configure Service Agent Access.
For an independent perspective, see Carlos E. Perez, Structuring Agents, Skills, and MCPs: Best Practices from Anthropic (May 6, 2026). This is commentary, not an official Anthropic publication.