What agent configs should compile to in 2026

If you're picking which per-agent files to emit from AGENTS.md, this is the rubric for May 2026: emit to legacy single-file targets, leave per-folder systems alone, and treat native AGENTS.md support as the absence of an emit target rather than a different one.

The rest of this post explains why, with the current-agent table and the per-target decisions Terso ships with as defaults.

The four shapes of agent config in 2026

Every popular agent today reads one of four shapes:

Native AGENTS.md — the agent looks at the root file directly. No emit target needed.
Root-level single file — the agent reads one specific file at a specific path (CLAUDE.md, .cursorrules, .github/copilot-instructions.md). Emit target candidate.
Per-path glob rules — the agent reads a directory of small files, each scoped to a glob. Not an emit target — these are authored where they apply.
Embedded in client config — the agent reads instructions from a field inside its config JSON/TOML (systemMessage, mcp_servers, etc.). Not really compilable; refer the user at AGENTS.md from inside the config.

Today's agents, mapped to shapes

Agent	Native AGENTS.md?	Root single file	Per-path rules	Embedded
OpenAI Codex CLI	✅ primary	—	—	—
GitHub Copilot	✅ (since Aug 2025)	`.github/copilot-instructions.md`	`.github/instructions/*.instructions.md`	—
Cursor	✅	`.cursorrules`	`.cursor/rules/*.mdc`	—
Claude Code	imports	`CLAUDE.md` (root + subdirs)	subdirectory `CLAUDE.md` files	—
Google Jules / Gemini	✅	—	—	—
Factory / Amp / Windsurf	✅	—	—	—
Zed	✅	—	—	—
Aider	reads via `--read`	any markdown file you point at	—	—
Continue	—	—	—	`systemMessage` in `.continue/config.json`

(RooCode shut down its agent product on May 15, 2026 and is no longer on the list. The table will keep shifting; the rubric below is the durable part.)

Two observations from the table:

Most new agents only support native AGENTS.md. Factory, Amp, Windsurf, Zed, Jules — for these, there's nothing to emit. Your AGENTS.md is the input.
The agents with the largest installed base support the legacy root-level file too. Copilot, Cursor, Claude Code. That's where the emit work pays off — installed bases that already have hand-edited files.

The compile rubric

For each agent your team uses, ask three questions:

Does it read AGENTS.md natively? If yes and there's no legacy file in your repo, you're done. Don't add an emit target.
Does the repo already have the agent's legacy single-file instructions? If yes, emit to it. The contributors who edit that file expect it to be there; ripping it out before native support is universal is a regression.
Does the agent have per-path rules you want to use? Author them in place. Don't compile.

Applied to terso-cli's defaults:

Target	Emit by default	Reason
`CLAUDE.md`	✅	Claude Code has Opus 4.7 subagents; legacy single-file is widely deployed.
`.cursorrules`	✅	Cursor reads it; many repos rely on it for Composer + Cloud Agents.
`.github/copilot-instructions.md`	✅	Copilot code review + coding agent consume it.
`.cursor/rules/*.mdc`	❌	Authored in place.
`.github/instructions/*.instructions.md`	❌	Authored in place.
Subdirectory `CLAUDE.md`	❌	Authored in place.
Continue `config.json`	❌	Reference `AGENTS.md` from inside the config; don't generate JSON.
Aider	❌	Use `--read AGENTS.md` in your `.aider.conf.yml`.

What gets lost when you hand-maintain instead

Three concrete failure modes the "compile" model removes:

The reviewer can't tell which file is canonical. When CLAUDE.md, .cursorrules, and .github/copilot-instructions.md all have hand edits, code review can't trust any of them. With one source and a marker comment, the canonical file is unambiguous.
PRs that change rules don't fail on inconsistency. A contributor edits .cursorrules, doesn't touch the others. CI doesn't catch it. Three months later, the team realizes Cursor and Claude Code have been working from different specs. emit --check would have flagged the PR.
Adding a new agent costs nothing. When AGENTS.md is the source, supporting a new client is either "the client reads AGENTS.md natively, done" or "the client reads a legacy file, add an emit target." Either way, you don't rewrite content; you add a renderer.

What about per-folder rules?

They compose with AGENTS.md. Repo-wide truths stay in the root file and get compiled; folder-scoped guidance lives in .cursor/rules/, .github/instructions/, or subdirectory CLAUDE.md files and is authored where it applies. Compile what's central; keep local rules local.

(For the longer argument, see the next config boundary is the folder, not the repo.)

A short forward bet

The Agentic AI Foundation's AGENTS.md v1.1 proposal is converging on formal layering + discovery semantics. When that lands, the compile-vs-author boundary may move — for example, a future spec revision might bless a "canonical scoped rules" file format that would compile. Until then, the rubric above is the right one: legacy single files compile, per-folder rules don't, native AGENTS.md support means no emit target at all.

This will age — the table will, especially. Re-verify before acting on it; the rubric beneath the table is the durable part.