Field Notes: AI — December 2025

The month in one paragraph

December 2025 was the agent-ops month. Coding agents became less like clever IDE features and more like billable, auditable, policy-shaped infrastructure. The practical questions changed: what did the agent do, which tools did it call, what did it cost, which identity did it use, what artifact did it produce, and can the run be replayed later?

Actual field update

• Observability became mandatory: run traces, transcripts, step logs, and tool-call histories became real operating artifacts.
• Billing became architectural: usage-based pricing and agent-heavy workflows made cost attribution part of system design.
• Service identities mattered: agents needed scoped credentials, not broad human credentials.
• Sandboxing became baseline: isolated execution environments became the only sane way to allow broad tool use.

Robustness check

Strong claim: agent systems need replayable traces and cost accounting.

Strong claim: sandboxing and scoped credentials are core runtime requirements.

Moderate claim: specific vendor release names matter less than the common operational convergence.

Agentic design pattern change

The durable pattern became:

run id
→ step id
→ tool call id
→ input/output snapshot
→ cost/latency
→ approval event
→ artifact
→ review decision

If you cannot replay the run, you do not have an agent system. You have a stochastic side-effect machine.

Fallout

• Agent transcripts became compliance material.
• Cost per accepted artifact became more useful than raw token spend.
• Tool-call logs became debugging and governance surfaces.
• Organizations needed agent service accounts and least-privilege execution.

What builders should copy

• Store every tool call with arguments, result hash, latency, cost, and approval state.
• Track cost per completed task, not only model-token cost.
• Give agents scoped service accounts instead of borrowed user authority.
• Treat transcripts as first-class artifacts.

Resource sources

• OpenAI Agents SDK: https://developers.openai.com/api/docs/guides/agents
• LangGraph overview: https://docs.langchain.com/oss/python/langgraph/overview
• MCP Authorization: https://modelcontextprotocol.io/specification/2025-06-18/basic/authorization
• AIDev dataset of agentic PRs: https://arxiv.org/abs/2602.09185

Operator math (TeX)

$$S_{\text{unit}} = \frac{C_{\text{agent-month}}} {N_{\text{accepted-artifacts}}}$$