Where is the boundary?
Name the files, systems, and decisions the agent may touch. “The codebase” is not a boundary.
Field note / 01May 2026
The new shape of software work
The era of supervising code has arrived.
Agentic coding is quickly changing the bottleneck: less typing, more briefing, checking, and deciding what deserves to ship. The good news is that a small team can start deliberately.
Read the field guideBrief better✳Inspect the diff✳Keep a human in the loop✳Make the rollback boring✳Brief better✳Inspect the diff✳Keep a human in the loop✳Make the rollback boring✳
Latest note
Dispatch 06 · Independent test
10 min read
The practical question is no longer “can an agent write it?”
It is: how do you give an agent enough context to be useful, while retaining enough visibility to trust the result?
This is why agentic coding is a serious topic now. In its 2026 trends report, Anthropic describes a shift from writing code toward orchestrating agents that write code—alongside the unresolved work of oversight, quality, and security. Google Cloud frames a related move from individual tasks toward whole systems. Those are product announcements and research perspectives, not a guarantee of your team’s outcome. But they point to a useful operating change: the skill is becoming system direction.
Thesis
The leverage is real only when the review loop is real.
02 / A working model
Treat the agent as a capable, very fast new hire.
Not magic. Not an intern. Not an autonomous department. Give it an intelligible job, a bounded workspace, and a feedback loop that exposes its work.
What a useful loop looks likeNot to scale
01 / Context02 / Execution03 / Evidence04 / Decision
03 / The field guide
Three questions before the first run.
What counts as evidence?
Ask for test output, a diff, a failure note, and the assumptions the agent made—not just a confident summary.
Who holds the stop button?
A useful pilot has an accountable reviewer, a clear escalation route, and a way to roll back.
The point of a pilot
Don’t prove that AI is clever.
Prove that your process improves.
Choose one repetitive workflow. Record the baseline. Make the agent show its work. Let a responsible person say “no.” That is enough to learn something valuable.
Write the first brief04 / Start here
Put the idea under a little pressure.
01 / PLAN
Turn an idea into a bounded brief
Define the job, constraints, handoff, and expected evidence before selecting a tool.
02 / MODELMake assumptions visible
Use the estimator as a conversation starter—not a forecast you can outsource responsibility to.
03 / EXPLORECompare the building blocks
Use the early directory to orient yourself, then go to original documentation and test the fit.
04 / PILOTCreate a reviewable first run
A small, supervised workflow beats a giant promise you cannot inspect.
Source desk
Read the original material.
This page is a perspective built from current first-party reports. Source claims are linked; conclusions and the pilot framework are The Model Note editorial guidance.
01Anthropic2026 Agentic Coding Trends ReportFrames the shift as moving from writing code to orchestrating agents that write code, while stressing oversight, quality, and security.202602Google CloudAI agent trends 2026 reportHighlights the move from one-off tasks toward systems and emphasizes the role of team capability.202603GoogleA new era for AI SearchDescribes agentic capabilities in Search and the growing expectation that AI can act across longer, contextual tasks.19 May 2026
Editorial status: draft for human review • No product testing, commercial relationship, or performance outcome is represented here.