Enterprises deployed agents with real tools—CRM, code, browser—under tracing, permissions, and continuous eval.
After AutoGPT hype, 2024 saw serious agents: reliable function calling, sandboxes, human-in-the-loop, and per-task success metrics.
Vendors (OpenAI, Anthropic, Google, startups) offered frameworks. The bottleneck became evaluation and tool design, not only the LLM.
Productive agents foreshadowed general computer use and 2025–2026 autonomous software flows.
An SoK paper submitted to arXiv on September 11, 2026 by Mia, Wu, Uluagac and Amini argues attacks can succeed at an agent’s planning, memory or tool layers while the final response looks clean. TechTimes placed it on September 15 beside Amodei’s swarm warnings; it urges intermediate logging, least privilege and stopping reliance on output-only metrics.
On 22 September 2026 Teradata expanded its Tera assistant with a Context Engine, execution Harness and reusable Agent Skills for governed data engineering and science; availability planned for Q4 2026, with claims of fewer tokens than Claude Code on SWE-bench Pro.
On 22 September 2026 at Atlanta’s Transformation Excellence Summit, SAP integrated Signavio, LeanIX, WalkMe and Cloud ALM with an AI Agent Hub, an AI Act/NIST Governance Assistant and production agent monitoring.