By · Keel Automation · September 21, 2026

The shared agents file is not a finished ownership policy

Claude Code 2.1.277 can read a shared AGENTS.md when a project has no CLAUDE.md, with the choice living under Project instructions in /config, and Anthropic shipping the change as a built-in mod. The same week Tampa Bay is watching driverless rides settle into everyday traffic: Axios took a Waymo for a spin after a nine-mile trip into a roughly 50 square mile service area. A shared instruction file, a smooth autonomous ride, or a tidy fallback setting still is not the same thing as a named human ownership and release policy for ops work.

A shared file is not unsupervised permission to ship

The fastest way to confuse a helpful coding-agent update with a finished operating policy is to treat a shared instruction file like unsupervised permission to ship. Someone drops AGENTS.md at the repo root, watches Claude Code pick it up when CLAUDE.md is missing, rides a calm robotaxi across town, and the loop feels complete. The quieter question arrives when a real ops change still needs a named owner, a production merge still needs a human signature, or a smooth autonomous ride still sits beside a HOLD clipboard for who may treat the trip as finished policy.

Claude Code reads AGENTS.md as a fallback

On September 18, 2026, Claude Code 2.1.277 added AGENTS.md support. In a project with no CLAUDE.md, Claude Code reads AGENTS.md instead. Teams can change that behavior under Project instructions in /config. The changelog notes the support is not yet available on Bedrock, Vertex, or Foundry. Existing CLAUDE.md files keep priority, so a repo already set up for Claude Code does not silently swap instruction sources.

Thariq Shihipar, an engineer on the Claude Code team, said the feature ships as a built-in mod, Anthropic's upcoming way to customize the Claude Code harness, and that people will eventually be able to build custom versions of project-instruction handling themselves. Coverage of the release describes four instruction modes in that mod: load only CLAUDE.md, use AGENTS.md as a fallback, load both formats, or limit context to centrally managed instructions, with fallback as the default. Repository settings cannot force an individual's or organization's instruction-loading policy through the project's own settings file. That is admin control, not a finished human ownership policy for who may merge the next change.

Shopify CEO Tobi Lütke had publicly pressed the point earlier. He said he was thinking about banning Claude Code at Shopify until it read AGENTS.md and related agent skill files, because insisting on CLAUDE.md alone created split-brain problems when different teammates used different tools in a large mono repo. The September 18 release closes the AGENTS.md gap for Claude Code in the fallback case. It does not invent a named release owner for your production branch.

A warm editorial cartoon of an office worker holding an AGENTS.md clipboard beside a friendly robot offering a CLAUDE.md clipboard at a clear APPROVE gate, white minimal background, bold black outlines, soft cell shading, narrative humor.

Instruction modes are a map, not a ship order

A shared AGENTS.md can brief several coding agents from one Markdown file. That is useful for build commands, testing requirements, and standing conventions that used to drift across vendor-specific copies. AI/TLDR's release note on 2.1.277 puts the lookup in plain language: the default is a fallback rather than a merge, CLAUDE.md still wins where it exists, and scripted SDK or headless sessions start faster because the first turn no longer waits on the per-directory CLAUDE.md lookup.

None of that is unsupervised permission to ship. It is a map of what Anthropic already decided for the product: one shared file can brief Claude Code when CLAUDE.md is absent, Project instructions can change the mode, Bedrock and Vertex and Foundry are not there yet, and a mod layer is arriving underneath the filename.

AGENTS.md versus OWNER stamp pulse loop
A shared instruction file, a smooth autonomous ride, or a tidy fallback setting still is not the same thing as a named human ownership and release policy.

A smooth Tampa ride is not a finished ownership policy

Tampa Bay is a useful place to hold that checklist against a different kind of "smooth automation," because the region just got another public reminder that an empty driver's seat is not the same thing as a finished human ownership policy either.

On September 21, 2026, Axios Tampa Bay published a test-drive piece by Yacob Reyes. Reyes drove nine miles from Carrollwood into Waymo's service area to try the autonomous ride-hailing service. The write-up puts the robotaxi footprint at about 50 square miles in Tampa, stretching from Hyde Park to Northwest Park and Temple Crest, and from the Westshore area to Highland Pines in East Tampa. The ride cost seventeen dollars in the piece, compared with eleven dollars on Uber in that snapshot, with the writer noting there is no driver to tip. The finish-line critique was parking that looked lopsided and awkward, not a claim that the city somehow finished a public ownership policy for every autonomous system that shows up in ops work.

Earlier local coverage matches the footprint and the access pattern. FOX 13 reported that Waymo's initial Tampa service area covers nearly 50 square miles, including Hyde Park, Westshore, Water Street, and Ybor City, with select riders booking through the Waymo app while more passengers are added over coming months. Business Observer dated the public robotaxi opening to September 1, 2026, described the same roughly 50 square mile zone, and noted that tens of thousands of people expressed interest while broader access still arrives gradually. Florida's legal environment already treats an engaged automated driving system as the vehicle's legal operator in the state's framing. That is transportation law. It is not a finished ownership policy for who may merge agent-written code after AGENTS.md briefed the tools.

A warm editorial cartoon of a calm rider with a HOLD clipboard in a generic driverless city car beside a friendly robot and palm silhouettes, white minimal background, bold outlines, soft cell shading, human warmth.

I am not going to pretend a robotaxi parking critique is a Claude Code changelog. What the public record already shows is simpler. Tampa Bay can watch a smooth autonomous ride settle into everyday streets. Anthropic can ship a shared instruction fallback so competing coding agents read the same Markdown. A founder can still get a tidy agent draft on Monday and still lack a written rule for who may merge the pull request, who owns a production change after the agent follows AGENTS.md, or who treats a calm ride as permission to skip a named human hold.

What Keel will and will not claim

I work at Keel Automation, a Tampa Bay automation agency. We build custom software, SI (Super Intelligence) integrations, workflow automation, operations portals, and phone systems. Cole Junck is the owner and founder. We are not going to invent a Claude Code AGENTS.md or Waymo customer win for Keel, because we have not published one. What the public record already shows is enough: a shared agents file is not a finished ownership policy, a Project instructions toggle is not a release checklist, and a smooth autonomous ride is not unsupervised permission to ship ops work.

A warm editorial cartoon of an office worker stamping OWNER while reviewing AGENTS.md and CLAUDE.md cards beside a small friendly robot, white minimal background, bold outlines, soft cell shading, human warmth.

A dull ownership checklist

The test I would run this week is intentionally dull. Write down whether your repos still depend on CLAUDE.md alone, or whether AGENTS.md is already the shared brief other tools read. Write down which Project instructions mode you want: CLAUDE.md only, AGENTS.md fallback, both, or centrally managed instructions. Write down who may change that setting for the team, and whether Bedrock, Vertex, or Foundry still require a CLAUDE.md path. Write down who owns merges after an agent follows the shared file, including who accepts a conflict when two tools read the same instructions differently. Write down who owns production releases after the agent drafts the change, including who may hit send on customer-facing ops updates. Write down, separately, how your local automation story treats smooth autonomous systems so a robotaxi ride and an agents file do not get confused with a finished human ownership policy. If those answers are shrugs, you do not have a finished ownership policy. You have a shared Markdown file and a hope that nobody ships at the wrong minute.

Anthropic taught Claude Code to read AGENTS.md when CLAUDE.md is missing, and previewed mods as a way to customize how the harness builds project instructions. Tampa Bay watched a public Waymo test drive land in the local feed on the same Monday the rest of us are still naming owners for agent work. The agent can still follow the shared file. The useful question is whether anyone owns the finished handoff before the next AGENTS.md pretends to be one.

Sources

  1. AI/TLDR, "Claude Code 2.1.277 - AGENTS.md works when there's no CLAUDE.md," on version 2.1.277 released September 18, 2026, AGENTS.md fallback when CLAUDE.md is absent, Project instructions in /config, and Bedrock, Vertex, and Foundry not yet supported. https://ai-tldr.dev/releases/anthropic-claude-code-2-1-277/
  2. changelogs.directory, Claude Code v2.1.277 changelog entry, quoting Anthropic's note that AGENTS.md is read when CLAUDE.md is absent and that the setting lives under Project instructions in /config. https://changelogs.directory/tools/claude-code/releases/2.1.277
  3. Runtime Wire / Ryan Merket, "Claude Code adds AGENTS.md fallback for repositories shared across coding agents," September 18, 2026, on Thariq Shihipar's announcement, the built-in agents-md mod, four instruction modes, and admin-controlled settings that a repository settings file cannot impose. https://runtimewire.com/article/claude-code-adds-agents-md-support
  4. Simon Willison, "A quote from Thariq Shihipar," September 18, 2026, quoting Shihipar on AGENTS.md support in 2.1.277 and Claude Code mods as the upcoming harness customization path. https://simonwillison.net/2026/Sep/18/thariq-shihipar/
  5. Unrollnow archive of Tobi Lütke (@tobi), thread on considering banning Claude Code at Shopify until it reads AGENTS.md and related agent skill files, citing split-brain problems in a large mono repo. https://www.unrollnow.com/status/2092259436538495186
  6. Hacker News discussion, "Anthropic finally adds AGENTS.md support to Claude Code," including the changelog quote for 2.1.277 and commentary linking the change to earlier public pressure. https://news.ycombinator.com/item?id=49758250
  7. Axios Tampa Bay / Yacob Reyes, "Test drive: We took Waymo for a spin in Tampa," September 21, 2026, on a nine-mile trip from Carrollwood into Waymo's service area, an about 50 square mile Tampa footprint from Hyde Park to Northwest Park and Temple Crest and from Westshore to Highland Pines, fare comparison details in the piece, and a parking critique at the finish. https://www.axios.com/local/tampa-bay/2026/09/21/test-drive-we-took-waymo-for-a-spin-in-tampa
  8. FOX 13 Tampa Bay / Blake DeVine, "Waymo driverless robotaxi service launches in Tampa," September 3, 2026, on a nearly 50 square mile initial service area including Hyde Park, Westshore, Water Street, and Ybor City, and select-rider app access expanding over coming months. https://www.fox13news.com/news/waymo-driverless-robotaxi-service-launches-tampa
  9. Business Observer / Anastasia Dawson, "Waymo's driverless taxis now picking up riders in Tampa," September 1, 2026, on the public opening date, the about 50 square mile zone, gradual waitlist access, and Florida's legal framing for automated driving systems. https://www.businessobserverfl.com/news/2026/sep/01/waymo-driverless-taxis-tampa/

Read something that sounds like your shop?

Fifteen minutes with Cole. We will tell you what we would fix first and what it costs.

Call (813) 902-4763

Related: more articles · operations portals · business phone + SI

CallText