diff --git a/datadog.gemspec b/datadog.gemspec
index 97a39dbc8f..690cd8eb32 100644
--- a/datadog.gemspec
+++ b/datadog.gemspec
@@ -48,6 +48,7 @@ Gem::Specification.new do |spec|
       .select { |fn| File.file?(fn) } # We don't want directories, only files
       .reject { |fn| fn.end_with?('.so', '.bundle') } # Exclude local profiler binary artifacts
       .reject { |fn| fn.end_with?('skipped_reason.txt') } # Generated by profiler; should never be distributed
+      .reject { |fn| File.basename(fn) == 'AGENTS.md' } # Developer tooling; not useful to gem consumers
 
   spec.executables = ['ddprofrb']
   spec.require_paths = ['lib']
diff --git a/lib/datadog/open_feature/AGENTS.md b/lib/datadog/open_feature/AGENTS.md
new file mode 100644
index 0000000000..df4ea511fd
--- /dev/null
+++ b/lib/datadog/open_feature/AGENTS.md
@@ -0,0 +1,385 @@
+# OpenFeature coding guide
+
+This guide applies to contributors and their AI coding tools working under
+`lib/datadog/open_feature/`, `spec/datadog/open_feature/`, and `sig/datadog/open_feature/`.
+The goal is to keep review conversations focused on design and correctness rather than
+conventions. Each rule is grounded in a real finding from this codebase; the guide grows as
+new patterns emerge. When in doubt, follow the existing files in this directory as the
+reference.
+
+---
+
+## Naming
+
+Ruby values intention-revealing names. Every local variable, method, and parameter should say
+what it holds or does.
+
+**Variables: use full, descriptive names.**
+
+```ruby
+# bad
+flat = flatten_context(attrs)
+e    = entries.first
+s    = build_string(parts)
+
+# good
+flattened_context = flatten_context(attrs)
+first_entry       = entries.first
+encoded_string    = build_string(parts)
+```
+
+Single-letter names are acceptable only for established Ruby idioms:
+
+```ruby
+# fine: block index
+parts.each_with_index { |part, i| ... }
+
+# fine: rescue variable
+rescue => e
+  logger.debug(e.message)
+
+# fine: throwaway / ignored argument
+def handle(event, _context)
+```
+
+**Directories and modules must match.** A module named `FlagEvaluation` lives in
+`flag_evaluation/`, not `flagevaluation/`. Wire-level names (API endpoints, protocol fields)
+stay on the protocol side and do not leak into Ruby identifiers.
+
+**Use Ruby naming throughout, including comments and test descriptions.** Do not carry
+camelCase from other languages:
+
+```ruby
+# bad (in comment or RSpec description)
+# globalCap controls the upper bound
+
+# good
+# global_cap controls the upper bound
+```
+
+---
+
+## Comments
+
+Comment the *why*, not the *what*. Use the minimum words needed to explain the constraint,
+decision, or non-obvious behaviour. If a comment keeps growing, that is a sign the code
+itself needs to be clearer — extract a method or rename a variable instead.
+
+```ruby
+# bad: narrates obvious code across multiple lines
+# We iterate over the context attributes and increment the counter for each
+# one we process, stopping once we reach the maximum allowed field count.
+count = 0
+attrs.each { |k, v| count += 1 }
+
+# bad: one-liner that still just restates the code
+# increment the retry count
+retry_count += 1
+
+# good: one line, explains a non-obvious constraint
+# Cap at 256 to match the backend's field limit; extras are silently dropped.
+pruned[key] = value if pruned.size < MAX_CONTEXT_FIELDS
+```
+
+Do not restate a test description as a comment inside the example:
+
+```ruby
+# bad
+it "returns nil when the flag is missing" do
+  # returns nil when the flag is missing
+  expect(subject).to be_nil
+end
+
+# good
+it "returns nil when the flag is missing" do
+  expect(subject).to be_nil
+end
+```
+
+Do not include references that only make sense outside this repository (ticket IDs, "Node
+reference", "Python sibling"). The Ruby code stands alone.
+
+Use ASCII characters only. Do not use Unicode box-drawing dividers (`# ─── Section ───`);
+they are not used elsewhere in the codebase.
+
+---
+
+## Methods
+
+Keep methods small and single-purpose. Prefer many focused private methods over one large one.
+
+```ruby
+# bad: one method doing too much
+def process(event)
+  key = [event[:flag], event[:variant]].join(":")
+  return if @seen.include?(key)
+  @seen << key
+  payload = { flag: event[:flag], count: (@counts[key] || 0) + 1 }
+  @transport.send(payload)
+end
+
+# good: each step named and separated
+def process(event)
+  key = cache_key(event)
+  return if already_seen?(key)
+  record(key)
+  @transport.send(build_payload(key))
+end
+```
+
+Prefer `return unless condition` over `return nil unless condition` (Ruby implicitly returns
+nil):
+
+```ruby
+# bad
+return nil unless enabled?
+
+# good
+return unless enabled?
+```
+
+---
+
+## Error handling
+
+Rescue only what you intend to handle, and only as broadly as necessary.
+
+```ruby
+# bad: catches everything, hides real bugs
+begin
+  do_work
+rescue => e
+  logger.debug(e.message)
+end
+
+# good: deliberate, narrow, explained
+@queue.push(event, true)
+rescue ThreadError
+  # Queue full. Drop and count; backpressure is reported on the next flush.
+  @overflow_count += 1
+```
+
+A broad `rescue => e` at a product boundary (flag evaluation, trace pipeline) is acceptable
+when the intent is to never interrupt the caller. Scope it tightly and comment why.
+
+---
+
+## Types (RBS)
+
+Every file in `lib/datadog/open_feature/` has a matching signature in
+`sig/datadog/open_feature/`. Add or update the `.rbs` file in the same change as the Ruby
+file. CI enforces this.
+
+Do not suppress the type checker to make an error go away:
+
+```ruby
+# bad
+result = compute_value # steep:ignore
+
+# good: fix the type or add an explicit annotation
+result = compute_value #: String
+```
+
+If a suppression is genuinely unavoidable, add a comment explaining why and call it out in
+the PR description. Prefer `Type?` over `(nil | Type)`.
+
+---
+
+## Tests
+
+**Use verifying doubles.** `instance_double(RealClass)` raises when you stub a method that
+does not exist on the real class. String-name doubles and hand-rolled `Struct` fakes do not.
+
+```ruby
+# bad: passes even if Writer#enqueue is renamed or removed
+let(:writer) { double("Writer", enqueue: nil) }
+let(:writer) { Struct.new(:enqueue).new(nil) }
+
+# good: fails fast when the interface changes
+let(:writer) { instance_double(Datadog::OpenFeature::FlagEvaluation::Writer, enqueue: nil) }
+```
+
+**Tests must be independent.** Each example must set up its own state and clean up after
+itself. Do not rely on test ordering or on state left by a previous example. This matters
+especially when a global singleton (like `OpenFeature::API.instance`) is involved: reset it
+in `before`/`after` hooks for every example that touches it, and stub every message an object
+will receive.
+
+**Assert exact values when you know them.**
+
+```ruby
+# bad: passes even if count is wildly wrong
+expect(result.count).to be >= 1
+
+# good
+expect(result.count).to eq(3)
+```
+
+---
+
+## Concurrency and threads
+
+Background threads have a few mandatory properties in this codebase:
+
+- **Fork safety.** When a process forks (e.g. Puma spawning workers), all background threads
+  from the parent process die silently in the child — the child starts with no threads even
+  though the objects still exist. Use `Core::Workers::Async::Thread` with
+  `FORK_POLICY_RESTART`: it detects that the process was forked and restarts the thread
+  automatically on the next operation.
+
+- **Bounded queues.** Use `SizedQueue` with an explicit capacity rather than a plain `Array`
+  or `Queue`. An unbounded queue accepts events faster than they can be flushed; under
+  sustained load it grows until the process runs out of memory. A `SizedQueue` caps growth
+  and lets the producer handle the overflow explicitly.
+
+  ```ruby
+  # bad: grows without limit
+  @queue = Queue.new
+
+  # good: caps at 4096 entries; producer handles ThreadError on overflow
+  @queue = SizedQueue.new(4096)
+  ```
+
+- **Shutdown timeout.** Always call `join(timeout)` when stopping the thread. Without a
+  timeout, a stuck thread prevents the process from exiting.
+
+- **Non-blocking enqueue.** The hook runs on the caller's flag-evaluation thread. Never let
+  a full queue or a stopped worker stall that thread. Push non-blocking, catch the overflow,
+  and count the drop — do not raise or wait:
+
+  ```ruby
+  # bad: blocks the caller if the queue is full
+  @queue.push(event)
+
+  # good: drops and counts on overflow, never blocks the caller
+  @queue.push(event, true)
+  rescue ThreadError
+    @overflow_count += 1
+  ```
+
+---
+
+## Repo idioms
+
+These apply across the whole repository, not just this directory.
+
+**Time.** Use `Datadog::Core::Utils::Time.now` / `.get_time` instead of `Time.now`. The time
+provider is injectable (used by tests and Timecop integrations).
+
+```ruby
+# bad
+timestamp = Time.now.to_i
+
+# good
+timestamp = Core::Utils::Time.get_time
+```
+
+**Environment variables.** Read through `Datadog::Core::Environment::VariableHelpers` or the
+settings layer, never `ENV` directly. Run `rake local_config_map:generate` after adding a new
+environment variable.
+
+**`filter_map`.** Use `Core::Utils::EnumerableCompat.filter_map` instead of the native
+`filter_map`. Native `filter_map` requires Ruby 2.7+; this codebase supports Ruby 2.5 and
+2.6.
+
+**Settings and component wiring.** New configuration belongs in `Configuration::Settings`
+extended via `Core::Configuration::Settings.extend`. New components are built in
+`Component.build` and wired into `components.rb`. Follow the existing wiring pattern rather
+than inventing a new one.
+
+**Accessing private members.** `.send(:member)` is an accepted pattern here when no public
+accessor exists — the tracer uses it in several places. Prefer adding a real public accessor
+or seam when one is practical.
+
+---
+
+## Structure and PR hygiene
+
+- One class or module per file, matching the file name.
+- Keep PRs focused. Each PR should change one thing: a new feature, a refactor, or a bug
+  fix — not all three at once. Mixing concerns makes it hard to understand intent, hard to
+  revert, and slow to review.
+- No dead code. Dead code is any method, branch, or class that no current caller reaches.
+  Common forms: a method defined but never called anywhere; a `rescue` branch for an
+  exception the surrounding code cannot raise; an `if` branch whose condition is always
+  false; a "forward compatibility" path added for a future caller that does not exist yet.
+  Delete it — version control remembers it if it is ever needed again.
+
+  ```ruby
+  # dead: start_worker is defined but perform already starts the thread;
+  # nothing calls start_worker
+  def start_worker
+    perform
+  end
+
+  # dead branch: the SDK never dispatches :finally today, so this branch
+  # is never reached
+  def run_hook(stage, context)
+    case stage
+    when :before  then before(context)
+    when :finally then finally(context)  # no caller reaches here
+    end
+  end
+  ```
+
+- Always construct objects through their normal constructor. Do not use `.allocate` +
+  `instance_variable_set` to bypass `initialize` — in tests, benchmarks, or anywhere else.
+  If the constructor requires collaborators that are hard to supply, pass lightweight
+  test doubles through the constructor arguments instead:
+
+  ```ruby
+  # bad: bypasses initialize, breaks silently when the constructor changes
+  writer = Writer.allocate
+  writer.instance_variable_set(:@transport, NoopTransport.new)
+  writer.instance_variable_set(:@logger, logger)
+
+  # good: uses the real constructor, passes a lightweight stand-in
+  writer = Writer.new(transport: NoopTransport.new, logger: logger)
+  ```
+
+- **PR size.** If the resulting diff would exceed roughly 1000 lines of additions, stop and
+  warn the contributor before generating the code. Propose a split into smaller, stackable
+  PRs instead. Each PR should be reviewable on its own and mergeable independently. Get
+  agreement on the breakdown before generating any code.
+
+### PR description
+
+Use `.github/PULL_REQUEST_TEMPLATE.md` as the structure. Fill each section with one or two
+sentences — high-level intent, not a list of every changed file. The reviewer reads the diff
+for details; the description should answer *what* and *why*, and call out any non-obvious
+trade-off or deliberate decision.
+
+```
+**What does this PR do?**
+Adds an EVP writer that batches flag evaluation events and ships them to the Agent every 10s
+via the EVP proxy. The writer is fork-safe and drops events non-blocking on queue overflow.
+
+**Motivation:**
+Required by the FFE telemetry spec. Without this, flag evaluation counts never reach the backend.
+
+**Change log entry**
+Yes. Flag evaluation counts are now sent to Datadog via the EVP proxy.
+
+**Additional Notes:**
+Chose a canonical context key over MD5 so the encoding is auditable without a digest
+dependency. Retry on transport failure is deferred — the writer logs and moves on for now.
+
+**How to test the change?**
+Covered by the new aggregator and writer specs; integration verified against mock intake.
+```
+
+Respond to every review comment with either a fix or an explanation of
+why the change is not being made. Do not re-request review with open threads unresolved.
+
+---
+
+## Style
+
+Run `bundle exec rake standard:fix` before pushing. StandardRB is fixed and non-configurable
+by team convention.
+
+---
+
+**Tool note.** Claude Code reads `CLAUDE.md`, not `AGENTS.md`. If you use Claude Code,
+symlink this file: `ln -s AGENTS.md CLAUDE.md` inside this directory. Verify how your
+specific tool (Cursor, Copilot, Codex) discovers nested guide files.
diff --git a/spec/datadog/release_gem_spec.rb b/spec/datadog/release_gem_spec.rb
index 6e1fb00435..482ad77334 100644
--- a/spec/datadog/release_gem_spec.rb
+++ b/spec/datadog/release_gem_spec.rb
@@ -30,9 +30,9 @@
            |CONTRIBUTING.md
            |SECURITY.md
            |Gemfile
-           |AGENTS\.md
+           |(.*/)?AGENTS\.md
            |AGENTS_TODO\.md
-           |CLAUDE\.md
+           |(.*/)?CLAUDE\.md
            |(ruby|jruby)-\d+.\d+.gemfile
            |Rakefile
            |Matrixfile