Field Notes: AI — April 2026

The month in one paragraph

April 2026 was the durable-runtime month. The core agent loop was not the hard part anymore. The hard part was everything around it: state, compaction, tool permissions, workspace isolation, subagents, hooks, skills, session storage, review surfaces, and failure recovery. Long-running work across repositories and tools made the agent runtime more important than the chat interface.

Actual field update

• Long-running execution became the center: agents handled broader tasks with less synchronous supervision.
• Workspace topology mattered: worktrees, multi-root workspaces, and isolated environments became practical requirements.
• Context compaction became runtime engineering: long sessions needed controlled summarization, not random prompt stuffing.
• Subagents became useful where boundaries were real: parallel or specialist workers helped when their outputs were independently reviewable.

Robustness check

Strong claim: serious agent systems are simple loops wrapped in heavy infrastructure.

Strong claim: context management and workspace isolation are core architecture, not side features.

Weak claim to avoid: "subagents are always better." Subagents add cost and debugging complexity unless task boundaries are clear.

Agentic design pattern change

The pattern became:

outer runtime:
  policy, budget, state, checkpoints, approvals, observability

inner agent loop:
  observe, reason, call tool, update state, continue

In practical terms: graph outside, dynamic loop inside.

Fallout

• Run state needed persistence and recovery.
• Workspaces needed cheap branching and rollback.
• Context needed freshness, authority, and scope.
• Tool errors became state transitions, not just exceptions.

What builders should copy

• Keep an append-only run ledger.
• Snapshot workspace state before risky changes.
• Use worktrees/containers for speculative execution.
• Store context with source, freshness, authority, and scope.
• Treat tool errors as model-visible observations when recoverable.

Resource sources

• Claude Code design-space analysis: https://arxiv.org/abs/2604.14228
• LangGraph overview: https://docs.langchain.com/oss/python/langgraph/overview
• OpenAI Agents SDK: https://developers.openai.com/api/docs/guides/agents
• Anthropic — Building Effective Agents: https://www.anthropic.com/engineering/building-effective-agents

Operator math (TeX)

$$V_{\text{exec}} = \frac{N_{\text{tasks-completed}}} {H_{\text{human-supervision}}}$$