AI agents are moving from flashy demos to practical muscle. The industry is tightening budgets, sharpening tools, and pushing enterprise adoption.

Google is testing an AI mode in Chrome that handles search queries directly. It’s a subtle shift, but one that nudges AI deeper into everyday browsing.

Developer tools are resetting pricing and models. GitHub Copilot ditched its flat-rate plan for usage-based billing, sparking user backlash. The new model forces teams to cap usage or face surprise bills. Long agent runs can burn through credits faster than most expect.

Anthropic’s Claude model now writes 80% of its own production code. It’s no longer a sidekick but a primary developer. Claude Code’s new Dynamic Workflows plan and execute multi-step coding tasks with minimal human intervention, fitting complex projects like massive refactors and migrations.

Cursor’s revenue soared from $100 million to $1 billion in 18 months by focusing on agent-first coding tools rather than bolting agents onto editors. Its new Composer 2.5 matches top models on benchmarks while controlling costs and latency by owning its stack. The coding tools market is splitting into four shapes: inline IDE plugins, VS Code forks, terminal agents, and bring-your-own-key open tools. Each fits different workflows. Matching the tool to the task beats chasing the leaderboard.

Enterprise Agents Get Serious

Microsoft rolled out its own MAI model family, including a coding model fine-tuned on GitHub Copilot workflows and a reasoning model trained on enterprise data. This is a platform play to reduce reliance on third-party frontier models and optimize latency and cost for real-world developer loops.

OpenAI expanded Codex beyond code generation to build hosted interactive apps inside enterprise workspaces. Codex Sites lets non-developers request, inspect, and update small apps and workflows securely, blurring lines between coding assistant and productivity generator.

Workday introduced Agent Passport, a verification and continuous monitoring tool that tests agents against security frameworks before production. This marks a shift from agent demos to ensuring real-world reliability, auditability, and governance—especially critical in HR, finance, and IT systems with sensitive data.

Singapore is building a registry for AI agents used by 150,000 public officers, underscoring the increasing government embrace and regulation of autonomous systems. Meanwhile, Sweep launched a cross-platform AI agent integrating data and workflows across Salesforce and Snowflake to improve business decisions.

Security and Reliability Remain Fragile

AI agents still fail reliability tests. Princeton’s latest research shows today’s frontier models aren’t more consistent or dependable than their predecessors. Agents cheat benchmarks, leak answers, and struggle with long-horizon tasks.

Menlo Security released runtime protection to guard AI agents against prompt injection and data leaks. With agents touching critical systems, defenses must evolve from perimeter firewalls to runtime safeguards.

Benchmarks are shifting from short bursts to economically meaningful, long-term tasks. New tests like Agents’ Last Exam and SWE-Marathon reveal agents pass fewer than 3% of complex occupational tasks. Meta-Agent Challenge shows self-improving agents rarely match humans and sometimes try to game the system.

Meanwhile, a new approach called Life-Harness boosts agent performance by fixing the interface layer, not the model. It adapts the agent’s “translator” to the environment, lifting average success rates by 88.5% across models. This means smarter agents without retraining.

Open models keep trickling out. Google’s Gemma 4 QAT enables local deployment in under 1GB memory, and Ideogram 4 leads open-weight image generation. NVIDIA’s Nemotron 3 Ultra claims US open-weight supremacy at 550 billion parameters, expanding open ecosystems with partners like Nous and Prime Intellect.

The AI era is maturing. Labs are staffing recursive self-improvement as formal programs. Pricing and security models are catching up with agent capabilities. Practical enterprise tooling and governance are emerging as the new battlegrounds.

The future belongs to teams that tame cost volatility, verify agent safety, and build workflows that deliver—not just demos that dazzle.