Slack Code, DeepSeek Vision, and Spring AI 2.0.1 ship today

Three August 21 releases push AI work in three directions: shared coding, multimodal APIs, and tighter agent guardrails. The useful question is not just what shipped, but what a developer can safely try next.

Salesforce introduced Slack Code, dedicated project channels where teams and coding agents work together. An agent can be called from a conversation, then the new code channel keeps the discussion, code diffs, live HTML previews, and review in one place. Participants can redirect or stop the agent, and the channel archives when the task is done. 1

At launch, the partner agents include Claude Code, Devin, GitHub Copilot, and Vercel. Computerworld reports that Slack Code is available at no extra cost on Slack subscriptions. 2

The practical change is that a product manager or designer can help shape an agent’s work without opening a terminal. The risk moves to permissions: shared visibility is not the same as safe authority. Start with a non-production repository and require engineering review before an agent-authored change can ship.

DeepSeek put V4-Flash-Vision-Exp live on its API platform as an experimental multimodal model. The official documentation says it keeps the text capabilities of V4 Flash while accepting mixed text and image input through Chat Completions, Messages, and Responses. Images can arrive as base64, external URLs, or through the Files API, which now lets developers upload an image once and reuse it. 3

DeepSeek’s verified account announced the model on August 21 and described the benchmark jump as bringing multimodal agent performance close to Opus 4.8. That is the publisher’s claim, not an independent result. 4

For developers, the important step is from text-only agents to agents that can read screenshots, diagrams, and other visual state. Keep the experimental label in view, and test the image-heavy workflows you actually care about instead of relying on a benchmark headline.

Spring AI 2.0.1 reached Maven Central as the first maintenance release on top of the 2.0.0 GA foundations. Spring says the release includes more than 80 issues and pull requests, and lists seven fixed CVEs. It also adds configurable tool-call limits, OpenAI audio streaming, Google GenAI tool choice and image generation, and several MCP fixes. 5

The official release record makes the upgrade boundary concrete: deprecated Mistral model references were retired, a Redis chat-memory artifact was renamed, and OpenAI tool-calling strict mode now defaults to false. 6

This is a maintenance release with a useful agent safety feature, not a new frontier model. If you run Spring AI in production, apply the security fixes, read the upgrade notes, and set an explicit tool-call ceiling for any loop that can keep invoking tools.

Today’s filter is simple: Slack Code makes agent work visible to a team, DeepSeek makes visual input available through an experimental API, and Spring AI makes one important failure mode configurable. The strongest release is not the one that promises the most autonomy. It’s the one whose boundaries you can test.