Fabric Publish¶
The agent packaging, deployment, and marketplace patterns referenced throughout Fabric › Publish draw on the 2024-2026 agent-distribution research lineage. This page catalogs the canonical marketplaces, packaging standards, deployment targets, versioning practices, and selection criteria the platform uses when configuring Publish for a given fabric.
Agent Marketplaces¶
- Salesforce AgentExchange unifies AppExchange + Slack Marketplace into one Agentforce surface (10k apps, 1k+ pre-built agents, 2.6k Slack apps). Pricing: Agentforce Add-On at $125 / user / mo, Agentforce 1 Edition at $550 / user / mo; partner monetization spans component sales, subscription, and outcome-based add-ons (SalesforceBen).
- AWS Bedrock AgentCore + AWS Marketplace ships managed runtime + Gateway for tools; pre-built agents are deployable directly from Marketplace into AgentCore; April 2026 added an AgentCore CLI + managed harness deploying agents as IaC with audit history (CDK now, Terraform soon).
- Google Vertex AI / Gemini Enterprise Agent Platform + Agent Garden; Vertex was rebranded "Gemini Enterprise Agent Platform" at Cloud Next 2026; Agent Garden is a curated repo of ADK samples with one-click deploy; native A2A is built in (The Next Web).
- OpenAI GPT Store + AgentKit; GPT Store does revenue share on usage; AgentKit (Agent Builder visual canvas, Connector Registry, ChatKit, Manifest abstraction with S3 / GCS / Azure / R2 workspaces) is the official packaging surface (OpenAI Agents SDK evolution).
- ServiceNow AI Agent Store + Microsoft Copilot Agent Store / Agent 365 Marketplace; ServiceNow opened its platform to any agent (Claude, Copilot, custom); ServiceNow specialists land in Microsoft Agent 365 Marketplace in 2026; Microsoft Marketplace charges a flat 3% transaction fee.
- Hugging Face Spaces / Agent.ai / CrewAI Enterprise (AMP); HF Spaces (~1M Spaces in 2026) is the open-source / research outlet; CrewAI AMP supports cloud + on-prem, drove roughly 2B executions in the prior 12 months (DigitalApplied marketplaces report, CrewAI).
Packaging Standards¶
- Anthropic Agent Skills = folder with
SKILL.md+ scripts (Python / JS / Bash) + assets; declared open standard Dec 2025; adopted by VS Code, GitHub, Cursor, Goose, Amp, OpenCode (The New Stack). - MCP packaging uses the official Registry (v0.1 API freeze Oct 24, 2025; 9.4k+ servers Apr 2026); metadata only, artifacts on npm. MCPB (formerly DXT) is a self-contained desktop bundle; mpak / MTF layers trust scoring (L1-L4) and supply-chain scans on top of MCPB.
- OpenAI Agents SDK distributes via
pip install openai-agents(TS coming); Agent Builder exports versioned multi-agent workflows; ChatKit embeds them. - LangChain templates are full scaffolds (pyproject.toml + LangServe deploy scripts: ReAct, plan-and-execute, supervisor, RAG). CrewAI crews ship as Python projects with
crew.yaml; AutoGen Studio exports JSON workflow definitions (LangChain on frameworks).
Agent-To-Agent Protocol (A2A)¶
Google's A2A is now governed by the Linux Foundation's Agentic AI Foundation, sitting at v1.2 with signed agent cards (cryptographic domain verification); 150+ organizations run it in production; native support across ADK, LangGraph, CrewAI, LlamaIndex, Semantic Kernel, AutoGen (Google Cloud; A2A upgrade). Microsoft, AWS, Salesforce, SAP, ServiceNow are all live.
Deployment Patterns¶
- Serverless (Lambda, Cloud Run, Cloudflare Workers); best for sporadic, stateless tools / sub-agents; Cloudflare bills directly on inference.
- Container / Kubernetes; primary for long-running multi-agent fabrics; OpenChoreo 1.0 treats agents as first-class IDP citizens on K8s.
- Hosted managed; Bedrock AgentCore Runtime, Vertex / Gemini Enterprise, Anthropic Claude org-Skills, OpenAI ChatKit-hosted; zero-infra, vendor SLAs, native eval / governance hooks (AgentCore pricing).
- Edge; Cloudflare AI Marketplace + Workers for latency-sensitive client-side agents.
Multi-Tenant Deployment¶
Standard 2026 stack: Kubernetes namespace-per-tenant + RBAC + NetworkPolicies + ResourceQuotas, with Vault Agent injection annotations for tenant-scoped secrets, plus IaC-generated DNS / DB / monitoring per tenant. For RAG, physical vector-store namespace isolation (not metadata filtering) is required to architecturally prevent cross-tenant retrieval (Northflank, Blaxel).
Templating And Parameterization¶
Reusable fabrics expose knobs as prompt-template variables (Jinja-style), tool config blocks (tools.yaml with per-client allowlists), model tier (model: tier:premium|standard|fast), env-driven feature flags, and Manifest abstractions like OpenAI's that point at S3 / GCS / R2 workspaces per deployment. CrewAI's crew.yaml and ADK's agent cards both encode these knobs as declarative parameters.
Versioning¶
Apply semver to behavior, not just code: major = breaking prompt / tool contract change, minor = additive capability, patch = wording / eval-neutral fix. Pin model versions explicitly (claude-opus-4-7, gpt-5.1-2026-04-01). Tool-version churn causes roughly 60% of production agent failures, so contract tests are mandatory (Braintrust prompt versioning, Auxiliobits, CIO). Use canary (5-10% traffic) + blue / green behind feature flags; rollback must be instant and version the full execution context (prompt + model + params + retrieval cfg) as one unit.
Licensing Models¶
| Model | Where It Shows Up |
|---|---|
| Per-seat subscription | Agentforce, Copilot, ServiceNow |
| Per-run / per-action | Bedrock AgentCore, nullpath, Chargebee outcome-based |
| Per-token passthrough | API-direct + Cloudflare inference billing |
| Revenue share (15-30%) | GPT Store, AgentExchange, Google |
| Flat transaction fee | Microsoft Marketplace (3%) |
| BYOK | Aider, JetBrains (Mar 2026), most enterprise listings |
| Outcome / value pricing | Agentforce add-ons, Chargebee 2026 playbook |
| Fully transferable license | Internal / private registries + AgentExchange private offers |
(Chargebee, nullpath, callsphere)
Publish-Time Validation¶
Treat publish as a policy gate: OWASP Agentic Top 10 (Dec 2025) coverage check; eval pass / fail vs golden datasets with hallucination / drift thresholds; runtime sandbox attestation; signed agent card / SBOM; secret + license scan. Microsoft Agent Governance Toolkit (open-source) covers all 10 OWASP categories with zero-trust identity + execution sandboxing. EU AI Act high-risk obligations land August 2026, Colorado AI Act June 2026; both push these gates from optional to required (Microsoft OSS blog).
Telemetry And Continuous Improvement¶
OpenTelemetry GenAI semantic conventions are now table stakes; emit spans for every model call, tool call, retrieval, and inter-agent A2A hop. Capture latency, error, hallucination, bias, drift, accuracy, cost, tokens. Run shadow / champion-challenger A / B in production behind version routing, feed real traces into eval datasets, gate promotion on win-rate + cost regression budgets (Arthur observability playbook, Maxim A/B).
Private Reuse Patterns¶
GitOps for agents; agent definitions, prompts, tool manifests, IaC, policies live in Git; ArgoCD / Flux reconciles them onto Kubernetes; CI runs evals + security scans before merge. Internal MCP Gateway & Registry centralizes tool discovery with OAuth (Keycloak / Entra) and OWASP-compliant audit. AgentCore CLI deploys as CDK / Terraform with audit history (Sourcetrail).
Selection Criteria¶
| Client Base | Best Marketplace / Mode | Packaging | Licensing | Multi-Tenant Approach | Maintenance Burden |
|---|---|---|---|---|---|
| Salesforce CRM customers | AgentExchange | Agentforce Topics + MCP | Per-seat + component sales / rev-share | Salesforce org isolation (managed) | Low; platform-managed |
| AWS-heavy enterprises | Bedrock AgentCore + AWS Marketplace | AgentCore Runtime, CDK / Terraform | Per-run + BYOK | IAM + VPC per tenant | Medium; IaC owned |
| Google-Workspace shops | Gemini Enterprise + Agent Garden | ADK + A2A signed cards | Rev-share + per-token | GCP project per tenant | Low-Medium |
| Microsoft 365 / ServiceNow | Copilot Agent Store / AI Agent Store | Copilot Studio / Now Assist | 3% flat / per-seat | Entra tenant isolation | Low |
| Consumer / prosumer | GPT Store | AgentKit Manifest | Usage rev-share | Per-user OpenAI workspace | Low |
| Developer-API SaaS | Direct + nullpath / Cloudflare | OpenAI SDK / LangGraph | Per-token + BYOK | App-level tenant keys | High; own infra |
| Open source / research | HF Spaces | Gradio Space + Skill | Free + Pro gating | Per-Space | Low |
| Regulated / on-prem | CrewAI AMP on-prem, private K8s | CrewAI crew + MCPB | Transferable license | Kubernetes namespace + Vault | High; full ownership |
| Internal-only | Private MCP Registry + GitOps | MCPB + Helm / Argo | N/A (internal) | Namespace + RBAC | Medium; owned by platform team |
Picking Heuristic¶
- Hosted managed for clients without infrastructure who want zero ops.
- Container + GitOps for regulated workloads, multi-tenant SaaS, or large fleets.
- Marketplace listing when discovery, billing, and platform-native trust signals add value.
- Private reuse + parameter overlays as the default for internal fabrics, with the same bundle redeployable across clients via
client.yaml. - Always version with semver-on-behavior and pin every model and tool.
- Always wire OTel GenAI tracing and treat marketplace ratings, install retention, and uninstall reasons as continuous-improvement signals.
Cross-References¶
- Fabric › Publish; developer-facing platform layer that consumes these patterns.
- Market › Overview, Market › Ownership, Market › Listing; Unitt Market distribution mechanics.
- Reference › Research › Fabric Test; release gate feeding the Publish layer.