Team-standardizing AI coding agents: getting five developers to identical behavior

The individual case is solved: a developer who cares writes an
AGENTS.md and gets better agent behavior. The team case is where it
falls apart, five developers, two or three different agents, and a
codebase where every PR review re-litigates conventions the agents
never shared.

Why individual configs fail teams

1. The conventions live in heads, not files. Each developer's agent behaves like their habits, including the habits you wish they'd drop.
2. Configs disagree silently. Developer A's CLAUDE.md says table tests, developer B's says behavioral tests. Both agents look reasonable in isolation; the codebase gets both styles.
3. Review becomes the enforcement point. Every conventions gap surfaces as review comments, the most expensive place to fix them, paid by the person with the least context (the reviewer).

The standardization ladder

Level 1: One committed baseline. A single AGENTS.md in the repo,
committed like any code. Non-negotiable content: command table, error
and testing conventions, commit format, safety rules. This alone
removes most of the noise, because every agent in the repo reads at
least this (Claude Code via import; others via symlink, generation, or
the AGENTS.md convention).

Level 2: Tool layers derived, not independent. Cursor rules,
Claude-specific files, and the Copilot digest are projections of the
baseline, not separate documents with separate opinions. If a rule
exists only in a projection, it is either promoted to the baseline or
it is not a rule. This is the level that prevents the two-configs-
disagreeing failure.

Level 3: Behavioral skills shared. The procedures, how this team
does TDD, how bugs get investigated, what "done" means, live as
importable skill files alongside the baseline. New team members (and
their agents) inherit the workflow, not just the style.

Level 4: Validation in CI. Config integrity is checked like code:
required files present, budgets intact, frontmatter parses, no
placeholder text. A broken rule file that silently stops loading is a
conventions regression; treat it like one.

The politics (the actual hard part)

• Write the baseline in one sitting, together, fast. Two hours, one document, everyone in the room. The baseline that ships beats the one that's perfect.
• The command table is sacred. Whatever else is contested, exact commands are testable facts. Fix them and move on.
• Scope creep dies at the 300-line ceiling. When the baseline is full, the team must remove something to add something. That constraint is the feature.
• Personal preferences stay personal. Things that do not affect the shared codebase (theme, editor bindings, user-global memory files) stay out of the repo config entirely.

What a team kit looks like

The team tier of our AgentConfig Studio pack is exactly levels 1–4
packaged: twelve stack kits with derived tool layers, eight shared
workflow skills, and the validator for CI, with a license covering the
whole organization. But the ladder above works with hand-rolled files
too, the structure matters more than the source.

If you'd rather not assemble this by hand: AgentConfig Studio on Gumroad ships this as version-pinned, validator-tested kits for 12 stacks. The complete Next.js/TypeScript kit is free (MIT) if you want to inspect the structure first.