Enterprise agents go job-ready as model orchestration and cheap open weights…

Summary

Today's ten signals cluster around agents moving from toolkits to shipped products and around the model layer beneath them getting cheaper and more composable. OpenAI, Salesforce, and Zscaler each pushed agents deeper into governed enterprise workflows—analytics, CRM roles, and the SOC—while Sakana AI reframed 'pick a model' as 'pick an orchestrator' and DeepSeek kept cost pressure on with MIT-licensed open weights. Underneath, GitHub, Vercel, and Cloudflare hardened the operational plumbing: sandbox governance, connector permissions, free web grounding, and per-turn observability for voice agents.

Key themes

• Enterprise agents ship as job-ready products, not frameworks: OpenAI's ChatGPT Work Data agent brings natural-language BI over warehouses like Snowflake and BigQuery, Salesforce launched seven named Agentforce roles for sales, service, IT, and HR, and Zscaler's Agentic SOC applies autonomous triage and containment—each betting that grounding in existing systems of record and permissions is the differentiator.
• The model layer gets more composable and cheaper: Sakana AI's Fugu Ultra v2 and Fugu Max deliver a multi-agent orchestrator behind one OpenAI-compatible endpoint, competing with routing layers and single frontier models, while DeepSeek-V4.1-Flash pairs MIT open weights and a 1M-token context with steep off-peak pricing ($0.15/$0.60 per million tokens).
• Agent infrastructure and governance mature: enterprise-managed sandbox controls for GitHub Copilot in JetBrains, connector-management permissions in Vercel Connect, a bigger and more widely available Vercel Sandbox, free Tako live web search on Vercel AI Gateway, and turn-level metrics for Cloudflare voice agents all point to operability and control becoming table stakes.

Notable items

• OpenAI adds a Data agent to ChatGPT Work (built on GPT-5.6) for natural-language analytics and shareable dashboards across Snowflake, BigQuery, Databricks, Redshift, ClickHouse, MongoDB, and Datadog—high impact.
• Salesforce launches seven job-ready Agentforce agents (Casey, Paige, Carter, Hunter, Marshall, Piper, Fin) on Customer 360 ahead of Dreamforce; six are GA now, outbound-sales agent Hunter is in pilot with GA targeted for November 2026—high impact.
• Sakana AI releases Fugu Ultra v2 (capability-first) and Fugu Max (cost-first), an orchestrated pool of models behind one endpoint claiming five of eight benchmark wins in its own eval, including 48.3 vs Claude Opus 5's 27.3 on Chartography.
• DeepSeek-V4.1-Flash ships with MIT weights, a 552B-parameter sparse backbone (8B active per input token), a 1M-token context, and a 74.2 DeepSWE v1.1 score; it retires V4 Flash and absorbs V4 Pro traffic on September 14.
• Zscaler launches Agentic SOC, grounding Anthropic- and OpenAI-backed agents in its inline zero trust telemetry and a global decoy mesh for machine-speed triage, verdicting, and containment.
• Platform hardening from GitHub (enterprise-managed Copilot sandbox controls in JetBrains), Vercel (Connect connector permissions, Sandbox in all regions with 64 GB storage, free Tako web search on AI Gateway through September 30), and Cloudflare (turn-level voice-agent metrics in @cloudflare/voice v0.4.0).

Source coverage

Source rows used: 10