The Intake — Saturday, September 20, 2026

If you've been building agents that need to reach into Salesforce CRM data without maintaining a custom integration, Salesforce announced AIforce at Dreamforce on September 15. The architecture exposes Salesforce data, workflows, business logic, and governance to AI agents — no direct UI access required.

Three named implementations launched. Claudeforce connects Salesforce to Claude and is available in beta. It ships with 37 prebuilt sales skills covering prospecting and pipeline management. Slackforce brings Salesforce CRM context into Slack through Surfaces, Slackbot, and CRM features. Agentforce Coworker operates as an AI agent within Lightning. Salesforce reports 100,000 users activated in its first 35 days. The underlying architecture is Headless 360, introduced in March 2026. AIforce builds on its Headless Toolkit, which exposes MCPs, APIs, plugins, and skills. The AgentExchange marketplace launched alongside, with Anthropic, AWS, Google, OpenAI, DocuSign, and Vercel named as participants.

If you're building Claude-based agents that work with Salesforce CRM data, Claudeforce is now the beta path — 37 prebuilt sales skills scope the initial capability.

If your Gemini-based agent workflows need to handle large document sets or connect to third-party services like Jira, Salesforce, or GitHub, Google shipped the antigravity-preview-09-2026 harness on September 18. It adds two new capabilities.

The Files API streams large datasets, media files, and unstructured documents directly into a running agent's working context without custom preprocessing. The Credentials API stores third-party service credentials encrypted at rest and in transit. At runtime, credentials are accessed only via scoped ephemeral tokens at the egress proxy — the model never sees them. The harness is available in Google AI Studio and the Interactions API, including on the free tier. It runs on Gemini 3.8 Flash by default; the model is configurable per interaction. Google reports a 22% hit-rate improvement on long multi-step tasks and a 9% boost on complex coding workflows relative to the prior harness. Both figures are Google-reported; no independent verification has been published as of this writing.

The antigravity-preview-05-2026 harness is deprecated effective October 5. If you're running workflows on the prior harness, October 5 is the migration deadline — requests redirect automatically after that date.

If you've been evaluating open-weight models for agentic workloads and base model performance in multi-step loops has been the constraint, Abacus.AI released the Smaug line on September 10. The three open-weight models are fine-tuned for agentic loop tasks and available on Hugging Face.

The three models target different use cases. Smaug Flash is based on DeepSeek V4 Flash 0731. It scored 61.1 on LiveBench agentic coding — a 14.3-point gain over the base model. Smaug Mini is based on Qwen3.8 27B (27 billion parameters) and targets multimodal use cases and smaller reasoning workloads. Abacus.AI reports 82.0 on IFBench and 41.8 on AutomationBench. Smaug Agentic is based on Kimi K3. It scored 64.6 on LiveBench agentic coding and 94.1 on GPQA Diamond. All three are also available via Abacus.AI's RouteLLM API; enterprise VPC hosting is offered. Abacus.AI says the fine-tuning technique improves agentic loop performance 15–20% without increasing inference cost.

All benchmark figures are Abacus.AI-reported; no independent evaluation has been published as of September 20. If open-weight agentic loop performance is part of what you're evaluating, those numbers have no independent verification yet.

GitLab 19.4, released September 17, ships MCP Server Tools in public beta. If you've been connecting agent tooling to platform data — issues, merge requests, pipelines — through custom API wrappers, native MCP support is now available.

The public beta enables AI agents to connect to GitLab platform data and operations via the Model Context Protocol. The same release brings an MCP Server Admin Panel to general availability. The panel gives platform admins a centralized place to manage which MCP server integrations are active and govern their access grants. The same release ships per-user AI credit caps and Credits Usage Visibility at general availability.

If you're building agent workflows on GitLab, the MCP Server Tools beta is now the standard-protocol path for platform data access. Your GitLab admin can now manage all MCP server integrations from the Admin Panel — no scattered configuration files.

---