Field Notes: AI — January 2026

The month in one paragraph

January 2026 was the repeatability month. The strongest agentic systems started looking less like one-off prompting and more like standard engineering loops: task intake, context assembly, tool execution, artifact generation, review queue, and merge or rejection. The bottleneck moved from "can the model generate something?" to "can the organization repeatedly review, trust, and improve what it generates?"

Actual field update

• Repeatable loops became the unit: task → run → artifact → review → merge/reject.
• Schema-first integration hardened: structured tool inputs and outputs became the boundary between stochastic reasoning and deterministic systems.
• Review load became visible: human review moved from hidden labor to explicit capacity planning.
• Instruction files became common: CLAUDE.md / AGENTS.md-style project context became part of the runtime, but not a guaranteed win.

Robustness check

Strong claim: coding agents now leave measurable traces in commits and PRs.

Strong claim: review capacity is a real limiting factor.

Mixed claim: instruction files help. Empirical work suggests manifests can encode useful project rules, but vague or stale instructions can also degrade results.

Agentic design pattern change

The pattern became:

project manifest
→ task description
→ context pack
→ agent run
→ generated artifact
→ human review
→ feedback into future context

Context stopped being a prompt blob. It became project metadata plus execution state.

Fallout

• Teams needed instruction hygiene.
• Stale context became a new failure class.
• Review queues needed ownership and service-level expectations.
• Project conventions became machine-readable assets.

What builders should copy

• Version agent instructions with the repo.
• Separate durable rules from temporary task context.
• Track when an instruction last helped or harmed a run.
• Measure review load as a queueing problem.

Resource sources

• Agentic Much? Adoption of Coding Agents on GitHub: https://arxiv.org/abs/2601.18341
• AIDev — Studying AI Coding Agents on GitHub: https://arxiv.org/abs/2602.09185
• Agentic Coding Manifests / Claude.md study: https://arxiv.org/abs/2509.14744
• Anthropic — Building Effective Agents: https://www.anthropic.com/engineering/building-effective-agents

Operator math (TeX)

$$L_{\text{review}} = \lambda_{\text{incoming-runs}} \cdot \mathbb{E}[T_{\text{review}}]$$