AI Agents: Automation

Personal and operational automation has consolidated around parallel subagents with scoped tool access — six independent workers holding different contexts simultaneously — governed by hard guardrails (never send email autonomously, never make pricing decisions, default to "prep" not "dispatch" when uncertain). The compounding pattern is layered: an overnight inbox scan improves morning triage, better triage enables subagent dispatch, reliable dispatch makes time-blocking viable, and 36 hours of work compounds on itself. The newest automation layer packages repeated agent behavior itself: Browserbase distributes researched web-task playbooks, claude-smart captures local mistakes as reusable rules, and coding-agent best practices are evolving fast enough that teams need periodic protocol reassessment. Workflow pattern detection is now a first-class discipline: multi-source pattern analysis across Codex sessions, Memories, and Chronicle identifies candidates using a four-criteria filter (2+ occurrences, stable inputs/outputs, material improvement potential, not already covered), then packages them as the minimal viable form — Skills for reusable workflows, Custom subagents for bounded specialist tasks, Automations for scheduled monitoring. Planning itself has become an automation target: Sean Geng's plan-optimizer skill applies iterative self-scoring to keep the best plan each cycle and stop when improvements become noise. Evals-first development (define clear success criteria before starting, iterate until all evals pass while capturing learnings to a vault) rounds out the automation maturity stack.

The frontier is agent-first operations restructuring the org: Marcus Moretti runs Spiral at Every as a one-person team, replacing 60% of a PM's old week with strategy.md plus a /ce:product-pulse cron and a Now/Next/Later kanban (no sprints, standups, PRDs, or stakeholder updates), while Every's broader evidence suggests automation can expand the amount of human work by making expert competence cheaper and increasing demand. The unit of automation has shifted from the job to the cross-functional process, producing new roles — the "agent engineer" (internal-FDE wiring governed agents to Box/Salesforce/Workday) and matching "agent PM." OpenAI's "Lord Bottleneck" pattern shows the incremental build path: accelerate single tasks, chain the wins into one skill, then schedule it, and personify the system to make it approachable. Peter Yang's seven-criteria framework for an ideal personal agent (cross-tool, proactive, memory that "just gets you," multimodal, messenger-reachable, personable) sets a bar that no current agent — OpenClaw, Claude Code, or Codex — yet clears, confirming personal agents are a harder problem than engineering agents.