AI Weekly Malaysia

Dan McAteer argues that AI agents started working well around Christmas 2025 not because of model improvements alone, but because the 'agent harness'—the tools, context, memory, guardrails, and environment surrounding model weights—matured in tandem with model capabilities. His thesis is that models will progressively absorb harness functions into their weights, leaving engineers to build what is essentially a harness for managing human attention rather than model behavior.

Why: If you are building agent systems, this frames a concrete architectural decision: stop over-investing in elaborate scaffolding (custom tool-use wrappers, manual context compaction, hand-rolled guardrails) that frontier models will likely absorb into their weights within a generation or two. Instead, start designing for the human-attention layer—how operators monitor, intervene, and verify agent actions—as the durable part of your stack.

ABC Legal, a 1,100-person US legal document delivery company, moved from scattered employee-built Claude automations running on personal desktops to a governed fleet of 50+ agents using Claude Managed Agents—shared workspaces, single audit/billing surface, always-on cloud execution. CTO Brandon Fuller's team treats every agent as code (prompt + config in a repo) and reports ~50% cost reduction on some human tasks and ~310 daily active users across departments as of July 2026.

Why: If you are shipping internal AI agents, the architectural pattern here is concrete and worth copying: define agents as text+config in a version-controlled repo, move them off individual machines into a managed runtime with centralized billing and observability, and let non-engineers build while engineers govern. The specific numbers (50+ agents, 310/1100 employees active) give a rough benchmark for enterprise agent adoption depth.