Claude 4.5 Built Slack in 30 Hours Straight—Here's My Take After Testing

Claude Sonnet 4.5 is emerging as a practical “work-builder” model—especially for Excel and PowerPoint—backed by a track record of long, careful coding runs (including a reported 30-hour rebuild of Slack). The significance isn’t just raw capability; it’s the model’s apparent preference for taking time, checking its work, and producing outputs that are usable in real office workflows. That matters because the path to more autonomous software development depends less on demos and more on whether models can reliably operate inside the constraints of specific task environments. Early testing described here suggests the model performs well when prompts are specific and well-structured; vague or poorly framed instructions lead to outputs that are hard to build or even unreadable. In other words, the advantage comes from intent: if users can define the work clearly and place the model in a framework that matches the task, the system becomes “superpowered.”

That same theme—agents and models becoming operational at scale—shows up in Walmart’s deployment of a “super agent” spanning 200+ AI tools, with a reported 95% autofix rate on bugs. The takeaway is less about any single agent and more about orchestration: federated agent workflows can be stitched across complex development ecosystems today, delivering measurable value. The message for builders is direct: agents are already functioning inside large enterprises, and scaling them is primarily a matter of workflow design, intent definition, and orchestration rather than waiting for a future breakthrough.

OpenAI’s product moves reinforce the idea that model readiness and distribution are being treated as a combined strategy. ChatGPT Pulse and Sora are framed as new advertising surfaces, signaling a push toward ads as a major business line—potentially extending beyond consumer use into B2B placements. The launch pattern also hints at how OpenAI will roll out future models: assess readiness, then attach the model to a new surface quickly. Sora’s evolution from earlier “not quite ready” timing to a consumer product launch is presented as evidence of that approach.

AWS is also leaning into agent infrastructure with an “agent core” MCP server, positioning it as open-source runtime/gateway/identity/memory plumbing that helps developers build production-ready agents that can securely call external tools and maintain context across sessions. The strategic angle is cloud economics: open-source can help preserve AWS developer mindshare and revenue by making AWS the default place to build.

Microsoft Copilot is moving in the same direction—expanding beyond a single model by enabling Copilot to work with Claude models and supporting a multi-agent enterprise strategy. The underlying concern is distribution: Microsoft wants to keep Copilot as the front door for office productivity, even if that requires integrating competing models. That sets up a new competitive dynamic where customers may negotiate less vendor lock-in as major platforms offer model choice inside familiar interfaces.

Finally, Salesforce’s “Agentforce” push aims to bring natural-language coding into enterprise environments with security governance and compliance controls, connecting agents to Salesforce org data. The pitch is clear—reduce shadow IT and provide a governed “vibe coding” path—but the real test will be adoption: selling to CTOs is easier than persuading product managers, marketers, and CS leaders to switch if the tool doesn’t deliver day-to-day value. Across all these updates, the through-line is operationalization: models and agents are moving from novelty to integrated, governed work systems—if the prompts, workflows, and distribution channels are right.