diff --git a/AGENTS.md b/AGENTS.md
new file mode 100644
index 0000000..e9a7038
--- /dev/null
+++ b/AGENTS.md
@@ -0,0 +1,130 @@
+# AGENTS.md
+
+## Issue Discovery Standard
+
+When looking for new issues, do not rely only on casual code inspection. Prefer evidence-backed discovery methods and only file issues that are narrow, reproducible, and worth tracking.
+
+### Core rule
+
+A new issue should only be created if it is:
+- verified on the current codebase state
+- narrow in scope
+- reproducible with a concrete command, test, request flow, or minimal example
+- supported by actual observed behavior
+- easy for another engineer to validate and pick up
+
+### Preferred issue-finding strategies
+
+Use these methods before proposing new issues:
+
+1. Run code paths, not just static scans
+- Prefer executing small targeted checks over only reading code.
+- Use focused commands, test nodes, or minimal scripts to validate behavior.
+
+2. Re-run stale or suspicious test suites
+- Look for:
+  - `XPASS`
+  - `xfail`
+  - skipped tests that may now be stale
+  - stale or obsolete tests that may no longer match reality
+- Unexpected passes and stale expected failures are strong sources of real issues.
+
+3. Compare public API surfaces
+- Check package-level exports vs submodule exports.
+- Look for duplicated helper names with different behavior.
+- Validate that documented imports and actual imports behave consistently.
+
+4. Compare claims vs runnable behavior
+- Check README/docs statements such as:
+  - "supported"
+  - "in progress"
+  - "planned"
+  - "coming soon"
+- Verify whether the smallest runnable path actually matches the claim.
+
+5. Compare neighboring execution paths
+- Look for behavior that differs across:
+  - runtime vs test client
+  - native backend vs fallback backend
+  - sync vs async
+  - middleware vs no middleware
+  - top-level import vs direct submodule import
+
+6. Use differential testing mindset
+- When relevant, compare behavior to a reference implementation or expected contract.
+- For framework-style repos, compare against the documented compatibility target.
+
+7. Use property-based or fuzz-style thinking
+- Especially for:
+  - routers
+  - parsers
+  - protocol handlers
+  - validation layers
+  - path/header/query normalization
+- Prefer generated edge cases where practical.
+
+8. Look for issue-worthy drift
+- Documentation says one thing
+- tests assume another
+- runtime does a third
+- public exports expose a fourth
+- These mismatches often produce strong, narrow issues.
+
+### What counts as a strong issue
+
+A strong issue has:
+- one concrete problem
+- one reproducible path
+- actual observed failure or mismatch
+- clear expected behavior
+- nearby passing controls or guards
+- narrow, testable acceptance criteria
+- explicit non-goals if scope could sprawl
+
+### What does not count as a strong issue
+
+Do not file issues that are:
+- speculative
+- based only on code inspection when runtime proof is possible
+- broad umbrella complaints
+- duplicate broad trackers without narrowing them
+- warning-only cleanup unless there is a compelling reason
+- multiple unrelated findings bundled together
+
+### Required issue evidence
+
+Before creating an issue, gather:
+
+1. Exact repro
+- One concrete command, test node, API call, UI flow, or script.
+
+2. Observed result
+- The actual failure output, incorrect result, mismatch, or unexpected behavior.
+
+3. Expected result
+- What should happen instead.
+
+4. Nearby passing checks
+- One or two closely related tests/commands/behaviors that still pass.
+
+5. Narrow acceptance criteria
+- The repro passes.
+- The expected behavior is present.
+- Nearby guards remain green.
+
+6. Non-goals
+- Clarify what the issue is not asking for.
+
+### Filing rule
+
+If the issue cannot be reduced to:
+- one concrete problem
+- one reproducible path
+- one observed failure
+- one expected behavior
+- one or two nearby passing checks
+- one narrow fix target
+
+do not file it yet.
+
+Gather better evidence first.
diff --git a/src/tools.zig b/src/tools.zig
index d935b83..b1482d5 100644
--- a/src/tools.zig
+++ b/src/tools.zig
@@ -95,7 +95,7 @@ fn validateIssueDraft(alloc: std.mem.Allocator, title: []const u8, body: []const
     if (std.mem.trim(u8, body, " \t\n\r").len == 0) {
         return std.fmt.allocPrint(
             alloc,
-            "issue body is required and must follow the CONTRIBUTING.md / AGENT.md issue template. Missing body.\n\nRequired template:\n{s}",
+            "issue body is required and must follow the CONTRIBUTING.md / AGENTS.md issue template. Missing body.\n\nRequired template:\n{s}",
             .{ISSUE_TEMPLATE},
         ) catch null;
     }
@@ -119,7 +119,7 @@ fn validateIssueDraft(alloc: std.mem.Allocator, title: []const u8, body: []const
 
     return std.fmt.allocPrint(
         alloc,
-        "issue body does not satisfy CONTRIBUTING.md / AGENT.md issue requirements. Missing required section(s): {s}.\n\nRequired template:\n{s}",
+        "issue body does not satisfy CONTRIBUTING.md / AGENTS.md issue requirements. Missing required section(s): {s}.\n\nRequired template:\n{s}",
         .{ joined.items, ISSUE_TEMPLATE },
     ) catch null;
 }
@@ -281,8 +281,8 @@ pub const tools_list =
     \\{"name":"get_project_state","description":"Return all open issues grouped by status label, all open branches, and all open PRs. Use this to understand current project state before picking up work.","inputSchema":{"type":"object","properties":{},"required":[]}},
     \\{"name":"get_next_task","description":"Return the single highest-priority unblocked issue that has no open branch. Use this to decide what to work on next.","inputSchema":{"type":"object","properties":{},"required":[]}},
     \\{"name":"prioritize_issues","description":"Apply priority labels (priority:p0–p3) to a set of issues based on their dependency order. Sinks (no dependents) get p0; independent issues get p2.","inputSchema":{"type":"object","properties":{"issue_numbers":{"type":"array","items":{"type":"integer"},"description":"Issue numbers to prioritize"}},"required":["issue_numbers"]}},
-    \\{"name":"create_issue","description":"Create a single GitHub issue with title, body, labels, and optional milestone. Automatically applies status:backlog if no status label is provided. This tool enforces the issue requirements from CONTRIBUTING.md and AGENT.md: the body must include One-sentence problem, Exact repro, Observed result, Expected result, Nearby passing checks, Acceptance criteria, and Non-goals. If parent_issue is provided, create_issue only appends `Parent issue: #N` to the issue body for context; use link_issues for explicit dependency relationships.","inputSchema":{"type":"object","properties":{"title":{"type":"string"},"body":{"type":"string","description":"Required. Must follow the CONTRIBUTING.md / AGENT.md issue template with the required evidence sections."},"labels":{"type":"array","items":{"type":"string"}},"milestone":{"type":"string"},"parent_issue":{"type":"integer","description":"Optional parent issue number to append as `Parent issue: #N` in the issue body for context only. This does not create a durable GitHub subtask/dependency relationship; use link_issues for explicit links."}},"required":["title","body"]}},
-    \\{"name":"create_issues_batch","description":"Create multiple GitHub issues in one call. Issues are fired concurrently in batches of 5 with a 200ms collection window. Use this after decompose_feature to create all issues at once. Each issue body must satisfy the same CONTRIBUTING.md / AGENT.md evidence template enforced by create_issue.","inputSchema":{"type":"object","properties":{"issues":{"type":"array","items":{"type":"object","properties":{"title":{"type":"string"},"body":{"type":"string","description":"Required. Must follow the CONTRIBUTING.md / AGENT.md issue template with the required evidence sections."},"labels":{"type":"array","items":{"type":"string"}},"milestone":{"type":"string"}},"required":["title","body"]}}},"required":["issues"]}},
+    \\{"name":"create_issue","description":"Create a single GitHub issue with title, body, labels, and optional milestone. Automatically applies status:backlog if no status label is provided. This tool enforces the issue requirements from CONTRIBUTING.md and AGENTS.md: the body must include One-sentence problem, Exact repro, Observed result, Expected result, Nearby passing checks, Acceptance criteria, and Non-goals. If parent_issue is provided, create_issue only appends `Parent issue: #N` to the issue body for context; use link_issues for explicit dependency relationships.","inputSchema":{"type":"object","properties":{"title":{"type":"string"},"body":{"type":"string","description":"Required. Must follow the CONTRIBUTING.md / AGENTS.md issue template with the required evidence sections."},"labels":{"type":"array","items":{"type":"string"}},"milestone":{"type":"string"},"parent_issue":{"type":"integer","description":"Optional parent issue number to append as `Parent issue: #N` in the issue body for context only. This does not create a durable GitHub subtask/dependency relationship; use link_issues for explicit links."}},"required":["title","body"]}},
+    \\{"name":"create_issues_batch","description":"Create multiple GitHub issues in one call. Issues are fired concurrently in batches of 5 with a 200ms collection window. Use this after decompose_feature to create all issues at once. Each issue body must satisfy the same CONTRIBUTING.md / AGENTS.md evidence template enforced by create_issue.","inputSchema":{"type":"object","properties":{"issues":{"type":"array","items":{"type":"object","properties":{"title":{"type":"string"},"body":{"type":"string","description":"Required. Must follow the CONTRIBUTING.md / AGENTS.md issue template with the required evidence sections."},"labels":{"type":"array","items":{"type":"string"}},"milestone":{"type":"string"}},"required":["title","body"]}}},"required":["issues"]}},
     \\{"name":"update_issue","description":"Update an existing issue's title, body, or labels.","inputSchema":{"type":"object","properties":{"issue_number":{"type":"integer"},"title":{"type":"string"},"body":{"type":"string"},"add_labels":{"type":"array","items":{"type":"string"}},"remove_labels":{"type":"array","items":{"type":"string"}}},"required":["issue_number"]}},
     \\{"name":"close_issues_batch","description":"Close multiple issues at once. Each issue is marked status:done. Use this instead of calling close_issue N times.","inputSchema":{"type":"object","properties":{"issue_numbers":{"type":"array","items":{"type":"integer"},"description":"Issue numbers to close"},"pr_number":{"type":"integer","description":"PR number that resolves all these issues (optional)"}},"required":["issue_numbers"]}},
     \\{"name":"close_issue","description":"Close an issue and mark it status:done. Optionally reference the PR that resolved it.","inputSchema":{"type":"object","properties":{"issue_number":{"type":"integer"},"pr_number":{"type":"integer","description":"PR number that resolves this issue"}},"required":["issue_number"]}},
@@ -3122,9 +3122,9 @@ fn handleRunTask(
 }
 
 test "tools_list encodes evidence-backed issue filing guidance" {
-    try std.testing.expect(std.mem.indexOf(u8, tools_list, "This tool enforces the issue requirements from CONTRIBUTING.md and AGENT.md") != null);
-    try std.testing.expect(std.mem.indexOf(u8, tools_list, "Each issue body must satisfy the same CONTRIBUTING.md / AGENT.md evidence template enforced by create_issue") != null);
-    try std.testing.expect(std.mem.indexOf(u8, tools_list, "CONTRIBUTING.md and AGENT.md") != null);
+    try std.testing.expect(std.mem.indexOf(u8, tools_list, "This tool enforces the issue requirements from CONTRIBUTING.md and AGENTS.md") != null);
+    try std.testing.expect(std.mem.indexOf(u8, tools_list, "Each issue body must satisfy the same CONTRIBUTING.md / AGENTS.md evidence template enforced by create_issue") != null);
+    try std.testing.expect(std.mem.indexOf(u8, tools_list, "CONTRIBUTING.md and AGENTS.md") != null);
     try std.testing.expect(std.mem.indexOf(u8, tools_list, "run_explorer") != null);
     try std.testing.expect(std.mem.indexOf(u8, tools_list, "Maximum time for agent execution (default 300, max 600)") != null);
     try std.testing.expect(std.mem.indexOf(u8, tools_list, "Maximum total time for the full chain (default 300, max 600)") != null);