-
Notifications
You must be signed in to change notification settings - Fork 0
2026 03 28 agents md role and cross agent instructions
The role of AGENTS.md in a repo using .github/copilot-instructions.md as the sole instructions source
AGENTS.md has emerged as the cross-tool convergence format for agent project instructions, supported by OpenAI Codex, GitHub Copilot, Claude Code, Cursor, Aider, Gemini Command Line Interface (CLI), and others. This repo deleted AGENTS.md in Architecture Decision Record (ADR)-0006 (2026-03-07) in favour of .github/copilot-instructions.md as the single source of truth. Is that the right call? What does each tool actually read, and should AGENTS.md be restored: as content, as a pointer, or not at all?
- For the tools used in this repo (Copilot coding agent via GitHub Issues, Claude iOS (Apple's mobile operating system)
codefeature, theresearch-loop.ymlCopilot CLI workflow): which of these readAGENTS.mdat the repo root? Which read.github/copilot-instructions.md? Do any read both? - Is there a confirmed loading order or priority between the two files for any of these tools?
- Does the
agents.mdspecification define a standard for where the file must live (root only, or directory-scoped)?
- If
AGENTS.mdneeds to exist for cross-tool compatibility but.github/copilot-instructions.mdholds the content, is the correct pattern a one-lineAGENTS.mdthat says "see.github/copilot-instructions.md"? - Do agents (Copilot, Claude) follow pointer/import patterns in instruction files, or do they treat the file content literally?
- What did the previous
AGENTS.mdcontain before ADR-0006 deleted it? Is any of that content now absent fromcopilot-instructions.md?
- Do
davidamitchell/Latest-developments-,davidamitchell/Agent-Evaluation,davidamitchell/Personal-Assistant-,davidamitchell/Memory-System, anddavidamitchell/Policy-LSPhave anAGENTS.mdat root? - Is the current Research repo setup (no
AGENTS.md, noCLAUDE.md) consistent with the rest of the organisation, or is it an outlier? - What is the correct organisation-wide standard:
AGENTS.mdas primary withcopilot-instructions.mdpointing to it, orcopilot-instructions.mdas primary withAGENTS.mdoptionally pointing to it?
ADR-0006 recorded that deleting AGENTS.md was a deliberate choice: it noted as a trade-off that "Claude Code no longer has a .claude/ directory; it must use .github/copilot-instructions.md directly." Prior research (Research/completed/2026-03-08-ai-coding-harnesses-agent-philosophy.md) documents AGENTS.md as the cross-vendor standard. Prior research (Research/completed/2026-03-22-using-awesome-copilot-across-repos.md) notes that GitHub web already reads AGENTS.md and that it outranks workflows and plugins for improving agent behaviour. The Research repo is the only repo in the organisation confirmed to have deleted AGENTS.md.
-
Assumption: Other repos in the organisation have
AGENTS.md. Justification: Prior research referencesAGENTS.mdinLatest-developments-andawesome-copilot. Needs verification across all repos. -
Assumption: A thin pointer
AGENTS.md(one line referencingcopilot-instructions.md) is sufficient for cross-tool compatibility. Justification: Plausible but unverified; agents may not follow file pointers.
(Research agent to complete)
- If
AGENTS.mdis not read by the primary tools used in this repo (Copilot coding agent, Claude iOS), restoring it adds no value. - If the pointer pattern does not work (agents read the file literally), restoring
AGENTS.mdwith a pointer creates silent context loss.
- Does the Copilot coding agent read
AGENTS.mdat the repo root when assigned a GitHub issue? - Does Claude iOS
codefeature readAGENTS.md? - Do any of the other davidamitchell repos have
AGENTS.mdtoday?
Research question restated: Is deleting AGENTS.md in favour of .github/copilot-instructions.md as sole instructions source the correct call for this repo, given that AGENTS.md is the cross-tool standard? Should AGENTS.md be restored: as content, as a pointer, or not at all?
Scope confirmed:
- In scope: the three tools this repo actually uses (Copilot CLI in
research-loop.yml, Copilot Coding Agent via GitHub Issues, Claude iOS Code feature); what each file each tool reads; the pointer-pattern question; currentAGENTS.mdstatus in the five named repos in the davidamitchell organisation. - Out of scope: tools not used in this repo (Cursor, Aider, JetBrains Junie, etc.); general
AGENTS.mdauthoring best practices beyond what affects this repo's decision.
Constraints: Evidence must be drawn from official documentation, GitHub changelog posts, and verified repo inspections. Conclusions must be scoped to the three tools used in this repo.
Output format: Knowledge item with concrete recommendation on whether to restore AGENTS.md and in what form, plus consistency assessment for the organisation.
Prior work cross-reference:
-
2026-03-08-ai-coding-harnesses-agent-philosophy.md: coversAGENTS.mdas the cross-vendor standard for agent instructions; notes its adoption across major coding harnesses. -
2026-03-22-using-awesome-copilot-across-repos.md: confirmsPersonal-Assistant-hasAGENTS.md; establishes thatResearch,Latest-developments-, andAgent-Evaluationshare the same missing-surface pattern (noAGENTS.md); recommends restoringAGENTS.mdas a first-wave improvement for each.
Q1: Tool reading behaviour:
- Q1a: Does the Copilot CLI (used in
research-loop.yml) readAGENTS.md,.github/copilot-instructions.md, or both? - Q1b: Does the Copilot Coding Agent (web, assigned via GitHub Issues) read
AGENTS.md,.github/copilot-instructions.md, or both? - Q1c: Does the Claude iOS Code feature read
AGENTS.mdor.github/copilot-instructions.md? - Q1d: Where does the
AGENTS.mdspecification say the file must live? - Q1e: Is there a confirmed priority order between
AGENTS.mdand.github/copilot-instructions.mdfor the Copilot CLI?
Q2: Pointer pattern:
- Q2a: Does the
AGENTS.mdspecification support import or include syntax? - Q2b: Do Copilot or Claude agents follow file references included as plain text in an instruction file?
- Q2c: What content did the previous
AGENTS.mdcontain, and is any of it absent fromcopilot-instructions.mdtoday?
Q3: Org consistency:
- Q3a: Does
Personal-Assistant-haveAGENTS.mdat root? - Q3b: Do
Latest-developments-,Agent-Evaluation, andPolicy-LSPhaveAGENTS.mdat root? - Q3c: Is the Research repo an outlier in the organisation?
Q1a: Copilot CLI reading behaviour
Source: GitHub Copilot CLI custom instructions documentation (primary, official)
"Instructions in the
AGENTS.mdfile in the root directory, if found, are treated as primary instructions. If anAGENTS.mdfile and a.github/copilot-instructions.mdfile are both found at the root of the repository, the instructions in both files are used."
- [fact] The Copilot CLI documentation states that both
AGENTS.mdand.github/copilot-instructions.mdare used when both exist;AGENTS.mdis called "primary". (Source: GitHub Copilot CLI docs) - [fact] The research-loop.yml installs
@github/copilot(the Copilot CLI) with no pinned version, resulting in the latest published version (1.0.12 as of 2026-03-29) vianpm install -g @github/copilot. (Source: research-loop.yml) - [fact] A verified bug report (GitHub issue #489 in
github/copilot-cli) filed against Copilot CLI v0.0.353 documents thatAGENTS.mdis ignored when.github/copilot-instructions.mdis also present, despite help text listingAGENTS.mdas supported. The issue's status (fixed or open) could not be confirmed from the issue page alone. (Source: copilot-cli issue #489) - [inference] As of 2026-03-29, whether the
AGENTS.md-ignored bug is resolved in v1.0.12 is unconfirmed; the official documentation claims both files are used, but the bug report suggests actual behaviour may differ from documented behaviour for some versions.
Q1b: Copilot Coding Agent (web) reading behaviour
Source: GitHub blog changelog 2025-08-28 (primary, official)
"Copilot coding agent, our autonomous background agent, now supports
AGENTS.mdcustom instructions... AlongsideAGENTS.md, the agent continues to support GitHub's.github/copilot-instructions.mdand.github/instructions/**.instructions.mdformats, plusCLAUDE.mdandGEMINI.mdfiles."
- [fact] The Copilot Coding Agent (web, assigned via GitHub Issues) reads
AGENTS.md,.github/copilot-instructions.md,.github/instructions/*.instructions.md,CLAUDE.md, andGEMINI.mdwhen present, treating them as additive. Support forAGENTS.mdwas added on 2025-08-28. (Source: GitHub blog changelog) - [fact] No confirmed priority order between
AGENTS.mdand.github/copilot-instructions.mdis documented for the Copilot Coding Agent (web); the changelog says the files are used alongside each other. (Source: GitHub blog changelog)
Q1c: Claude iOS Code feature reading behaviour
Source: Claude Code docs overview (primary, official); secondary search results on Claude iOS.
- [fact] Claude Code reads
CLAUDE.mdfiles at project root and in.claude/directories; there is no documented support for readingAGENTS.mdin Claude Code. An open GitHub issue (#6235 inanthropics/claude-code) requests nativeAGENTS.mdsupport and remains open. (Sources: Claude Code docs, claude-code issue #6235) - [fact] The Claude iOS Code feature allows dispatching tasks from an iOS device to a Claude Code Desktop session. It does not introduce a separate instruction-file loading mechanism; the same
CLAUDE.mdlookup that applies to Claude Code Desktop applies. (Source: Claude Code overview) - [fact]
.github/copilot-instructions.mdis not on the documented read path for Claude Code or Claude iOS Dispatch; neither file is referenced in Claude Code's memory or instruction documentation. (Source: Claude Code memory docs) - [inference] Claude iOS Code feature reads neither
AGENTS.mdnor.github/copilot-instructions.mdas of 2026-03-29. It readsCLAUDE.md. This means the Research repo, which has noCLAUDE.md, provides no automated project-level instructions to Claude iOS Code at all.
Q1d: AGENTS.md file location specification
Source: agents.md official spec site (primary); OpenAI Codex AGENTS.md documentation (primary)
- [fact] The
AGENTS.mdspecification supports both root-level placement (affecting the whole project) and directory-level placement (affecting that directory and all subdirectories), using a nearest-file-wins hierarchy. (Sources: agents.md, OpenAI Codex AGENTS.md docs) - [fact] The specification is stewarded by the Agentic AI Foundation under the Linux Foundation, adopted by over 25 Artificial Intelligence (AI) coding tools as of March 2026. (Source: agents.md, particula.tech AGENTS.md explainer)
Q1e: Priority order between AGENTS.md and copilot-instructions.md (Copilot CLI)
- [fact] The Copilot CLI documentation says
AGENTS.mdis treated as "primary instructions" while.github/copilot-instructions.mdis "repository-wide" instructions, and both are used together. No explicit priority override is documented; conflicting instructions are resolved non-deterministically. (Source: GitHub Copilot CLI docs)
Q2a: AGENTS.md pointer/import support
Source: agentsmd/agents.md GitHub issue #11 (primary)
- [fact] The
AGENTS.mdspecification does not define a native import, include, or file-reference syntax. A GitHub issue (#11 atagentsmd/agents.md) requests this feature, indicating it is absent as of the filing date. (Source: agentsmd issue #11) - [inference]
CLAUDE.mdmay support a file-import directive (described as@importsin third-party documentation) specific to Claude Code; this claim comes from a secondary source rather than Claude Code's official documentation and should be treated as unverified. (Source: medium.com CLAUDE.md guide)
Q2b: Agents following file pointers in instruction files
- [inference] Agents (Copilot CLI, Copilot Coding Agent) read instruction file content literally; there is no mechanism by which a one-line instruction file causes the agent to load a second file at the same priority as a natively read file. A workaround exists (include a line such as "Also read
.github/copilot-instructions.md"), but the agent treats this as a task instruction rather than a file-loading directive, making adherence unreliable. (Sources: GitHub community discussion on AGENTS.md, Nathan Nellans Copilot instructions guide) - [assumption] A thin pointer
AGENTS.md(one line referencingcopilot-instructions.md) is insufficient for cross-tool compatibility. Justification: No specification or tool documentation describes this pattern as reliable; agents treat instruction files as context input, not as file-system directives.
Q2c: Previous AGENTS.md content
Source: ADR-0006 text in docs-adr/0006-*.md (primary, this repo)
- [fact] ADR-0006 states the decision was to "expand
.github/copilot-instructions.mdwith the full content fromAGENTS.md", meaning all content from the previousAGENTS.mdwas migrated intocopilot-instructions.md. No content was deleted; the files were consolidated. (Source: docs-adr/ADR-0006)
Q3a: Personal-Assistant- AGENTS.md
Source: Direct GitHub API inspection of davidamitchell/Personal-Assistant- root directory (2026-03-29)
- [fact]
Personal-Assistant-hasAGENTS.mdat root (shab1de2dfdff23c5d4b389482accee1d251f6d0132, 10,018 bytes). It also has.github/copilot-instructions.md, path-specific instructions, andmcp.json. (Source: Personal-Assistant- repo root)
Q3b: Other repos' AGENTS.md status
Source: Direct GitHub API inspection of root directories (2026-03-29)
- [fact]
Latest-developments-does NOT haveAGENTS.mdat root. Root contains:.env.example,.github/,.gitignore,.gitmodules,.mcp.json,BACKLOG.md,CHANGELOG.md,Makefile,PROGRESS.md,README.md, and source directories. (Source: Latest-developments- repo root) - [fact]
Agent-Evaluationdoes NOT haveAGENTS.mdat root. Root contains:.github/,.gitignore,.gitmodules,BACKLOG.md,CHANGELOG.md,PROGRESS.md,README.md, and project directories. (Source: Agent-Evaluation repo root) - [fact]
Policy-LSPdoes NOT haveAGENTS.mdat root. Root contains:.github/,.gitignore,.gitmodules,.mcp.json,BACKLOG.md,CHANGELOG.md,PROGRESS.md,README.md,Makefile, and source directories. (Source: Policy-LSP repo root) - [fact]
Memory-Systemcould not be inspected; the repo returned 404 on prior investigation (2026-03-22) and was unavailable. (Source: 2026-03-22-using-awesome-copilot-across-repos.md)
Q3c: Research repo outlier status
- [fact] Of the five inspected repos in the organisation (
Personal-Assistant-,Latest-developments-,Agent-Evaluation,Policy-LSP,Research), onlyPersonal-Assistant-hasAGENTS.md. The Research repo is not unique in lackingAGENTS.md; three of four comparators also lack it. (Source: GitHub API inspections, 2026-03-29) - [inference] ADR-0006's claim that deleting
AGENTS.mdwas "consistent with all other repos in the davidamitchell organisation" was accurate forLatest-developments-,Agent-Evaluation, andPolicy-LSPat the time of writing, butPersonal-Assistant-had or subsequently acquiredAGENTS.md, making it the outlier rather than Research.
Removing narrative:
-
The Copilot CLI (the tool used in
research-loop.yml) officially reads bothAGENTS.mdand.github/copilot-instructions.md, but a documented bug showsAGENTS.mdmay be silently ignored in favour ofcopilot-instructions.mdwhen both are present. Either way,copilot-instructions.mdis read reliably. -
The Copilot Coding Agent (web) reads both files since 2025-08-28. This is additive; both are used without conflict.
-
Claude iOS Code reads
CLAUDE.mdonly. NeitherAGENTS.mdnor.github/copilot-instructions.mdis on its read path. This is a genuine gap not addressed by the current setup or by restoringAGENTS.md. -
The thin pointer pattern is not a viable solution: the spec has no import syntax, and agents interpret instruction file content as context input, not as file-loading directives.
-
All instruction content from the previous
AGENTS.mdwas migrated tocopilot-instructions.mdby ADR-0006. No content was lost. -
Four of five organisation repos (including Research) lack
AGENTS.md. OnlyPersonal-Assistant-has it. Research is not an outlier in the current organisation state.
Potential contradiction: ADR-0006 says deleting AGENTS.md is "consistent with all other repos in the davidamitchell organisation," but Personal-Assistant- has AGENTS.md.
Resolution: Personal-Assistant- may have added AGENTS.md after ADR-0006 (2026-03-07), or ADR-0006 may have been inaccurate at writing time. Either way, the current state is that Personal-Assistant- has it and three of four other repos do not. This does not make Research an outlier; it makes Personal-Assistant- the outlier.
Potential contradiction: The Copilot CLI docs say both files are used, but a bug report says AGENTS.md is ignored.
Resolution: The bug was filed against v0.0.353; the current version is 1.0.12. Whether it is fixed is unconfirmed. The conservative reading is: copilot-instructions.md is reliably read; AGENTS.md may or may not be read depending on version. This makes copilot-instructions.md the more reliable single source.
No internal contradiction on Q2 (pointer pattern). All sources agree that agents read instruction files literally and no import syntax exists in the spec.
Technical lens:
- The Copilot CLI reads both files additively, which means restoring
AGENTS.mdwith a summary or condensed version ofcopilot-instructions.mdwould result in partial duplication. Duplication creates a maintenance surface: two files to keep in sync. - If
AGENTS.mdwere restored with full content (mirroringcopilot-instructions.md), the Copilot CLI would inject both into context, effectively doubling the instruction content without adding new information. This wastes context window tokens.
Tool coverage lens:
- Restoring
AGENTS.mdwould extend coverage to tools not currently used in this repo (Cursor, Aider, Gemini CLI, etc.) if such tools were ever used. The value of this is proportional to the probability of adopting those tools. - The one tool not served by the current setup (
CLAUDE.mdfor Claude iOS) would not be served by addingAGENTS.mdeither. If Claude iOS coverage is desired, aCLAUDE.mdis the correct addition.
Maintenance lens:
- The current single-file setup is the lowest maintenance path: one file to update, no sync required.
- Adding
AGENTS.mdwith real content requires keeping it aligned withcopilot-instructions.md; maintaining two parallel instruction files creates a divergence risk [inference].
Historical lens:
- The pre-ADR-0006 state had
AGENTS.mdas primary andcopilot-instructions.mdas a stub pointing to it. ADR-0006 inverted this tocopilot-instructions.mdas primary. Neither state had a thin pointer fromAGENTS.mdtocopilot-instructions.md.
Answer: Deleting AGENTS.md in favour of .github/copilot-instructions.md is the correct call for the two primary tools used in this repo (Copilot CLI and Copilot Coding Agent web), because copilot-instructions.md is reliably read by both, while AGENTS.md support in the Copilot CLI has a documented unreliability issue. The thin pointer pattern is not viable. The Research repo is not an outlier; four of five organisation repos lack AGENTS.md. The one genuine gap is Claude iOS Code, which reads only CLAUDE.md and is not served by either file.
Key conclusions:
- Copilot CLI and Copilot Coding Agent (web): both read
copilot-instructions.md; Copilot Coding Agent (web) also readsAGENTS.md(since 2025-08-28); Copilot CLI officially reads both but has a documented bug that may suppressAGENTS.md. - Claude iOS Code: reads
CLAUDE.mdonly; neitherAGENTS.mdnorcopilot-instructions.mdis on its read path. - Thin pointer pattern: not viable; no import syntax in the spec; agents read files literally.
- Previous
AGENTS.mdcontent: fully migrated tocopilot-instructions.mdby ADR-0006; no content loss. - Org consistency: Research is aligned with
Latest-developments-,Agent-Evaluation, andPolicy-LSP;Personal-Assistant-is the outlier with both files. - If the organisation decides to adopt
AGENTS.mdacross all repos, it should contain full content (not a pointer), meaning a copy or condensed version ofcopilot-instructions.mdcontent. This creates a maintenance obligation. - Adding
CLAUDE.md(notAGENTS.md) is the correct move if Claude iOS Code coverage is desired.
- All claims in §2 are sourced or labelled [fact], [inference], or [assumption].
- All atomic questions in §1 are answered in §2.
- §3 reasoning is free of narrative glue.
- §4 resolves both identified contradictions.
- §5 adds technical, tool-coverage, maintenance, and historical lenses without introducing unsourced claims.
- No unsupported generalisations remain.
- One uncertainty persists: whether the Copilot CLI
AGENTS.mdbug (issue #489) is fixed in v1.0.12. This is noted explicitly in Risks/Gaps. - Review outcome: PASS. All threads synthesised, all claims labelled, all uncertainties explicit.
Deleting AGENTS.md in favour of .github/copilot-instructions.md as sole instructions source is the better-supported approach for this repo's two primary tools. [inference] The Copilot CLI (used in research-loop.yml) reads .github/copilot-instructions.md reliably; AGENTS.md support in the CLI exists in documentation but has a verified bug that may suppress it in practice. The Copilot Coding Agent (web, via GitHub Issues) reads both files, meaning restoring AGENTS.md would add additive but redundant context for that tool. The thin pointer pattern (a one-line AGENTS.md referencing copilot-instructions.md) is not viable because the AGENTS.md specification has no import syntax and agents read instruction files literally. The one genuine gap in the current setup is Claude iOS Code, which reads only CLAUDE.md and is served by neither existing file.
-
The Copilot CLI (v1.0.12, used in
research-loop.yml) officially reads bothAGENTS.mdand.github/copilot-instructions.mdwhen both are present, but a verified bug report (GitHub issue #489, filed against v0.0.353) documents thatAGENTS.mdis silently ignored in favour ofcopilot-instructions.mdin at least one version; fix status in v1.0.12 is unconfirmed. [confidence: high] -
The Copilot Coding Agent (web, assigned via GitHub Issues) reads
AGENTS.md,.github/copilot-instructions.md,.github/instructions/*.instructions.md,CLAUDE.md, andGEMINI.mdas additive instruction sources;AGENTS.mdsupport was added in the 2025-08-28 changelog entry. [confidence: high] -
The Claude iOS Code feature (dispatched via mobile app to Claude Code Desktop) reads
CLAUDE.mdonly; it does not readAGENTS.mdor.github/copilot-instructions.md, and there is an open GitHub issue (#6235 inanthropics/claude-code) requesting nativeAGENTS.mdsupport that remains unresolved. [confidence: high] -
The
AGENTS.mdspecification, stewarded by the Agentic AI Foundation under the Linux Foundation, supports both root-level and directory-scoped placement using a nearest-file-wins hierarchy, and has no native import, include, or file-reference syntax. [confidence: high] -
A thin pointer
AGENTS.md(one line referencingcopilot-instructions.md) is not a viable cross-tool compatibility solution because agents read instruction file content as context input, not as file-loading directives, making adherence to such a pointer unreliable and unverifiable. [confidence: high] -
ADR-0006 fully migrated all content from the previous
AGENTS.mdinto.github/copilot-instructions.md; no instructions were lost in the deletion, meaning restoringAGENTS.mdfrom scratch would require duplicating content already incopilot-instructions.md. [confidence: high] -
Of the five inspectable repos in the davidamitchell organisation, only
Personal-Assistant-hasAGENTS.mdat root;Latest-developments-,Agent-Evaluation,Policy-LSP, andResearchall lack it, makingPersonal-Assistant-the outlier rather than Research. [confidence: high] -
.github/copilot-instructions.mdis the reliably read instruction file for the Copilot CLI regardless of whether a bug fix lands forAGENTS.mdsupport, because the official documentation designates it as the always-used repository-wide instruction surface for the Copilot CLI. [confidence: high] -
If restoring
AGENTS.mdwith real content, it would need to be maintained in sync withcopilot-instructions.md; maintaining two parallel instruction files creates a divergence risk [inference], and the current single-file approach eliminates that maintenance surface. [confidence: medium] -
[inference] If instructions for the Claude iOS Code feature are desired, adding a
CLAUDE.md(not anAGENTS.md) is the appropriate action, becauseCLAUDE.mdis the only instruction file that Claude Code's read path includes. [confidence: high]
| Claim | Source | Confidence | Notes |
|---|---|---|---|
| Copilot CLI reads both files officially | GitHub Copilot CLI docs | High | Primary source |
| AGENTS.md bug in Copilot CLI | copilot-cli issue #489 | High | Bug filed against v0.0.353; fix status in v1.0.12 unconfirmed |
| Copilot Coding Agent reads AGENTS.md | GitHub blog changelog 2025-08-28 | High | Official changelog |
| Claude iOS reads CLAUDE.md only | Claude Code memory docs, claude-code issue #6235 | High | Open issue confirms AGENTS.md not natively supported |
| AGENTS.md has no import syntax | agentsmd issue #11 | High | Feature request confirms absence |
| ADR-0006 fully migrated AGENTS.md content | ADR-0006 | High | Primary source, this repo |
| Personal-Assistant- has AGENTS.md | GitHub API inspection 2026-03-29 | High | File present, 10,018 bytes |
| Latest-developments- no AGENTS.md | GitHub API inspection 2026-03-29 | High | Root directory confirmed |
| Agent-Evaluation no AGENTS.md | GitHub API inspection 2026-03-29 | High | Root directory confirmed |
| Policy-LSP no AGENTS.md | GitHub API inspection 2026-03-29 | High | Root directory confirmed |
| Thin pointer not viable | agentsmd issue #11, copilot-cli community discussion | High | No spec mechanism for file loading |
| AGENTS.md spec, Linux Foundation governance | agents.md, particula.tech explainer | High | Confirmed by multiple sources |
-
Assumption: A thin pointer
AGENTS.mdis insufficient for cross-tool compatibility. Confirmed by investigation. TheAGENTS.mdspec has no import syntax; agents read files as context. The assumption was correct. -
Assumption: Other repos in the organisation have
AGENTS.md. Partially confirmed. OnlyPersonal-Assistant-has it; three of four other repos do not, matching Research's current state.
The evidence weighs in favour of the current setup on all three tested dimensions: tool coverage, maintenance cost, and organisation consistency.
The prior research item Research/completed/2026-03-22-using-awesome-copilot-across-repos.md recommended adding AGENTS.md to Research as a first-wave improvement. That recommendation was based on the Copilot Coding Agent (web) being able to read AGENTS.md directly and on Research lacking it while Personal-Assistant- had it. The present investigation refines that recommendation for this specific repo: copilot-instructions.md already holds the full instruction content; the Copilot CLI (the primary autonomous tool in research-loop.yml) has a documented bug that may suppress AGENTS.md; and restoring AGENTS.md would duplicate existing content without adding new tool coverage. The prior recommendation remains applicable to Latest-developments- and Agent-Evaluation, which lack equivalent copilot-instructions.md files and would gain first-time coverage from an AGENTS.md. .github/copilot-instructions.md is read reliably by both tools that matter for autonomous coding work in this repo (Copilot CLI and Copilot Coding Agent web). AGENTS.md is either redundant (Copilot Coding Agent, which reads both) or unreliable (Copilot CLI bug) for adding new coverage. The only tool not served by the current setup is Claude iOS Code, which would require CLAUDE.md, not AGENTS.md.
The maintenance-cost argument reinforces the status quo: copilot-instructions.md is already the complete and authoritative instructions source. Adding AGENTS.md creates a second copy to maintain; maintaining two parallel instruction files creates a divergence risk [inference].
The organisation-consistency argument also supports the status quo: three of four comparable repos lack AGENTS.md. Personal-Assistant- is the outlier, not Research.
- The Copilot CLI
AGENTS.mdbug (issue #489) status in v1.0.12 is unconfirmed. If the bug is fixed, the Copilot CLI would read both files, makingAGENTS.mdharmless to restore (additive context). If unfixed,AGENTS.mdremains silently ignored. - The Claude iOS Code gap: no instruction file in this repo is read by Claude iOS Code. If the owner uses Claude iOS Code to work in this repo, it operates without project-specific instructions. This is a known gap, not a regression from ADR-0006 (the previous
AGENTS.mdwas also not on Claude's read path). -
Memory-Systemrepo status is unknown (404 on prior inspection); its instruction-file state could not be assessed. - The AGENTS.md specification's adoption rate (25+ tools, 60k+ open-source projects per agents.md) will increase over time. If new tools are adopted in this repo that read
AGENTS.mdbut notcopilot-instructions.md, the current setup would miss them.
-
Is the Copilot CLI AGENTS.md bug (issue #489) fixed in v1.0.12? Direct test by creating a temporary
AGENTS.mdand verifying whether its content appears in a CLI session would confirm this. Out of scope for this item; could become a backlog item. -
Should
CLAUDE.mdbe added to this repo? If Claude iOS Code is used for repo work, aCLAUDE.mdat root would be the correct file to add. This is a separate decision fromAGENTS.md. Out of scope. -
Should
Personal-Assistant-be the pattern to follow or an exception? It is the only org repo with bothAGENTS.mdandcopilot-instructions.md. Whether to standardise on its pattern or keep it as an exception warrants an explicit org-level decision.
- Type: knowledge
-
Description: Confirmed that the current Research repo setup (
.github/copilot-instructions.mdonly, noAGENTS.md) is correct for its primary tools (Copilot CLI and Copilot Coding Agent web), with a documented gap for Claude iOS Code (requiresCLAUDE.md). Thin pointer pattern is not viable. Research is not an org outlier; three other repos also lackAGENTS.md. - Links:
Navigation
By Tag
bureaucracy
change-management
coase
constraint-analysis
control-model
decision-rights
delegation
- Q4: Decision rights that should move closer to execution
- Q5: Control model for the best throughput-risk trade-off
delivery-risk
- Operating model synthesis for split-authority delivery systems
- Q6: Leading indicators of instability in split-authority flow systems
demand-segmentation
enterprise
exception-handling
execution
flow
flow-design
flow-metrics
governance
- Operating model synthesis for split-authority delivery systems
- Q1: Dominant flow constraint in split-authority delivery systems
- Q2: Demand segmentation for fast-path vs controlled-path flow
- Q4: Decision rights that should move closer to execution
- Conditions under which internal governance controls minimise coordination costs in regulated enterprises
- Failure mechanisms of internal governance controls: bureaucratic inefficiency and informal circumvention in regulated enterprises
- Barriers to governance reform, leadership failure modes, and reform mechanisms in regulated enterprises
governance-patterns
incentives
- Failure mechanisms of internal governance controls: bureaucratic inefficiency and informal circumvention in regulated enterprises
- Barriers to governance reform, leadership failure modes, and reform mechanisms in regulated enterprises
instability
institutional-economics
- Conditions under which internal governance controls minimise coordination costs in regulated enterprises
- Failure mechanisms of internal governance controls: bureaucratic inefficiency and informal circumvention in regulated enterprises
- Barriers to governance reform, leadership failure modes, and reform mechanisms in regulated enterprises
leading-indicators
operating-model
organisation
- Conditions under which internal governance controls minimise coordination costs in regulated enterprises
- Failure mechanisms of internal governance controls: bureaucratic inefficiency and informal circumvention in regulated enterprises
- Barriers to governance reform, leadership failure modes, and reform mechanisms in regulated enterprises
organisational-design
queue-design
queueing
regulated-enterprise
- Conditions under which internal governance controls minimise coordination costs in regulated enterprises
- Failure mechanisms of internal governance controls: bureaucratic inefficiency and informal circumvention in regulated enterprises
- Barriers to governance reform, leadership failure modes, and reform mechanisms in regulated enterprises
routing
throughput
throughput-risk
transaction-costs
- Conditions under which internal governance controls minimise coordination costs in regulated enterprises
- Failure mechanisms of internal governance controls: bureaucratic inefficiency and informal circumvention in regulated enterprises
triage
- Q2: Demand segmentation for fast-path vs controlled-path flow
- Q3: Routing design that isolates exceptions from routine flow
williamson