diff --git a/AGENTS.md b/AGENTS.md
index fe7de8531d0..a45b3c2ebd3 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -56,6 +56,12 @@ docs/                     Developer documentation (see below)
 ./gradlew :path:to:module:test -PtestJvm=11  # Test on a specific JVM version
 ./gradlew spotlessApply                   # Auto-format code (google-java-format)
 ./gradlew spotlessCheck                   # Verify formatting
+./gradlew :path:to:module:forkedTest      # Run forked tests (separate JVM per class)
+./gradlew :path:to:module:latestDepTest   # Test against latest framework version
+./gradlew :path:to:module:muzzle          # Check instrumentation version safety
+./gradlew checkInstrumentationNaming      # Validate instrumentation module names
+./gradlew forbiddenApisMain               # Check forbidden API usage
+./gradlew checkAgentJarSize               # Verify agent jar < 32 MB
 ```
 
 ## Code conventions
@@ -67,6 +73,24 @@ docs/                     Developer documentation (see below)
 - **Forked tests**: Use `ForkedTest` suffix when tests need a separate JVM
 - **Flaky tests**: Annotate with `@Flaky` — they are skipped in CI by default
 
+### Instrumentation module naming (CI-enforced)
+- Module names **must end with a version number** (e.g., `couchbase-3.1`) or a configured suffix (`-common`, `-stubs`, `-iast`)
+- Module names **must contain their parent directory name** (e.g., `play-2.4` under `play/`)
+- Validated by: `./gradlew checkInstrumentationNaming`
+- Modules under `:dd-java-agent:instrumentation:datadog` are exempt (internal features)
+
+### InstrumenterModule separation
+- `InstrumenterModule` (the module descriptor) should be **separate from** `Instrumenter` (the bytecode advice). Recent refactoring systematically separated these.
+- Use `InstrumenterModule.Tracing` for tracing instrumentations, `InstrumenterModule.ContextTracking` for pure context propagation.
+- Use `@AppliesOn` to override the target system when an advice applies to a different product than its module.
+
+### Cleanup expectations
+- Remove unused imports
+- Remove duplicate condition checks
+- Remove unnecessary semicolons
+- Use `System.arraycopy()` instead of manual array copy loops
+- Use Java-style array declarations (`String[] args`) not C-style (`String args[]`)
+
 ## PR conventions
 
 - Title: imperative verb sentence describing user-visible change (e.g. "Fix span sampling rule parsing")
@@ -81,5 +105,108 @@ Code running in the agent's `premain` phase must **not** use:
 - `java.util.logging.*` — locks in log manager before app configures it
 - `java.nio.*` — triggers premature provider initialization
 - `javax.management.*` — causes class loading issues
+- `UUID.randomUUID()` — can trigger `java.util.logging` initialization via ACCP. Use `RandomUtils.randomUUID()` instead.
 
 See [docs/bootstrap_design_guidelines.md](docs/bootstrap_design_guidelines.md) for details and alternatives.
+
+## Performance guidelines (critical)
+
+The tracer runs inside every customer JVM. Overhead budget is extremely tight — every allocation, lock, and exception matters on the hot path.
+
+### Allocation avoidance
+- **Never use `Objects.hash()`** in hot paths — varargs creates an `Object[]` on every call. Use `HashingUtils.hash()` (fixed-arity overloads, zero allocation) or manually unroll the polynomial hash.
+- **Avoid primitive boxing** — use typed setters (`TagMap.set(key, int)`, `DDSpanContext.setMetric(key, long)`) instead of `Object`-accepting methods. The v0.4/v0.5 serializers were specifically refactored to eliminate boxing.
+- **Prefer `UTF8BytesString`** over `String` for span metadata (operation names, service names, resource names, tag keys). It lazily caches the UTF-8 byte representation, avoiding repeated `getBytes(UTF_8)` allocations during serialization.
+- **Reuse objects per-thread** when possible — `CoreTracer.ReusableSingleSpanBuilder` pools span builders per thread, providing major throughput gains in constrained heaps.
+- **Never throw exceptions for flow control** on hot paths — use pre-validation instead of catch-based parsing (e.g., fast-path format validation instead of catching `NumberFormatException`).
+- **Avoid `String.split()`, `String.replaceAll()`, `String.replaceFirst()`** — these compile regexes internally. Use `Strings.replace()` from `datadog.trace.util.Strings` or manual parsing.
+- **Use `StringBuilder` sparingly** — consider whether the string is actually needed or whether the computation can be done directly.
+
+### Use existing performance utilities
+- **Caching**: Use `DDCaches.newFixedSizeCache(capacity)` — NOT `ConcurrentHashMap` unless the key set is implicitly bounded. The `FixedSizeCache` family is open-addressing, lock-free, and bounded.
+- **Hashing**: `HashingUtils.hash(a, b, c)` — overloads for 1-5 args, plus primitive overloads for `int`, `long`, `boolean`, etc.
+- **Queues**: Use `Queues.newMpscArrayQueue()` / `Queues.newSpscArrayQueue()` from `datadog.common.queue` — VarHandle-based on Java 9+, JCTools on Java 8. Never use `LinkedBlockingQueue` or `ArrayBlockingQueue`.
+- **Random UUIDs**: Use `RandomUtils.randomUUID()` — NOT `UUID.randomUUID()` which can trigger `java.util.logging` initialization during premain.
+- **String operations**: `Strings.replace()`, `Strings.join()`, `Strings.truncate()` — regex-free alternatives.
+- **Atomic fields**: Prefer `AtomicLongFieldUpdater`/`AtomicReferenceFieldUpdater` on `volatile` fields over `AtomicLong`/`AtomicReference` objects — saves 16+ bytes per instance. Critical for frequently-allocated objects like `DDSpan`.
+
+### Serialization hot path
+- The v0.4/v0.5 trace serializers are performance-critical. Changes to `TraceMapperV0_4` or `TraceMapperV0_5` must be benchmarked.
+- Tag iteration uses `TagMap.EntryReader` to avoid boxing — use this API, not `getTag()` calls in tight loops.
+- UTF-8 encoding uses `SimpleUtf8Cache` (tag names, low cardinality) and `GenerationalUtf8Cache` (tag values, mixed cardinality) — both tolerate benign races for performance.
+
+### Benchmarking
+- Performance-sensitive changes **must include JMH benchmarks** in the appropriate `src/jmh/` directory.
+- Existing benchmarks live in `internal-api/src/jmh/`, `dd-trace-core/src/jmh/`, and `benchmark/`.
+- Run benchmarks: `./gradlew :module:jmh`
+- CI has a performance SLO gate — regressions in key benchmarks block merge.
+
+## Forbidden APIs and common pitfalls
+
+The build enforces forbidden API rules at compile time (`./gradlew forbiddenApisMain`). Key rules:
+
+| Forbidden | Use instead |
+|---|---|
+| `Objects.hash(Object...)` | `HashingUtils.hash()` (in hot paths) |
+| `String.split()` / `replaceAll()` / `replaceFirst()` | `Strings.replace()` or explicit `Pattern` |
+| `UUID.randomUUID()` | `RandomUtils.randomUUID()` |
+| `System.getenv()` / `System.getenv(String)` | `ConfigHelper.getenv()` |
+| `System.out` / `System.err` | SLF4J logging (shaded `datadog.slf4j`) |
+| `Class.forName(String)` | Framework-specific classloading utilities |
+| `ElementMatchers.named()` / `.hasSuperClass()` | `NameMatchers` / `HierarchyMatchers` equivalents |
+| `FixedSizeStripedLongCounter` | JDK `LongAdder` |
+| `Field.set()` and reflection setters | Avoid mutating final fields (JEP-500) |
+
+### Null safety
+Many recent bug fixes address NPEs in decorators, extractors, and context propagation. When writing advice or decorator code:
+- **Always null-check return values** from `AgentSpan.context()`, request/response getters, and header accessors.
+- **Guard decorator calls** — e.g., `if (path != null) span.setResourceName(path)`.
+- **Protect exception advice** from `NoClassDefFoundError` — the target class may not be loadable in all environments.
+
+### Resource management
+- **Close resources in reverse order** of opening.
+- **Streams wrapping other streams** must delegate `close()` to the underlying stream.
+
+## Thread safety and concurrency
+
+### General rules
+- **Prefer `volatile` over `AtomicInteger`/`AtomicReference`** for simple single-field CAS-free access.
+- **Use `AtomicFieldUpdater` patterns** when CAS is needed on frequently-allocated objects — avoids allocating a separate `Atomic*` wrapper.
+- **Tolerate benign race conditions** in caches — `FixedSizeCache` deliberately avoids synchronization; the worst case is a redundant computation. Do NOT add synchronization to cache code unless there is a correctness issue.
+- **Lock `selector.selectedKeys()`** before iterating — NIO selector key sets are NOT thread-safe.
+- **Reduce lock contention** — prefer lock-free structures (`VarHandle` queues, `ConcurrentHashMap`, `LongAdder`) over synchronized blocks.
+- **Avoid class loading under locks** — can cause deadlocks. Preload classes before entering critical sections.
+
+### Virtual threads
+- Virtual thread context tracking is supported. `ThreadUtils.isVirtualThread()` checks at runtime.
+- Thread-local reuse patterns (like `ReusableSingleSpanBuilder`) check for virtual threads and skip reuse if detected.
+
+## Dependency and build rules
+
+### Gradle conventions
+- **Use lazy APIs** — `tasks.register()` not `tasks.create()`, `tasks.named()` not `tasks.getByName()`, `configureEach {}` not `all {}`. The build has ~630 projects and ~33,000 tasks; eager configuration causes major slowdowns.
+- **Single GAV string** for dependencies: `implementation("group:artifact:version")` not the map syntax.
+- **Use the version catalog** (`gradle/libs.versions.toml`) for shared dependencies. Instrumentation modules declare framework versions directly.
+- **Instrumented framework deps must be `compileOnly`** — they must not leak into the agent jar.
+- **Script plugins (`gradle/*.gradle`) are deprecated** — use convention plugins in `buildSrc/` for new work.
+- **Agent jar size limit: 32 MB** — enforced by `checkAgentJarSize` task.
+
+### Shading / relocation
+- All third-party libraries bundled in the agent jar **must be relocated** (e.g., `org.slf4j` → `datadog.slf4j`, `okhttp3` → `datadog.okhttp3`).
+- OpenTracing must **never** be a direct dependency — enforced at build time.
+- See `dd-java-agent/build.gradle` `generalShadowJarConfig()` for the full relocation table.
+
+## API boundaries
+
+| Module | Visibility | Contents |
+|---|---|---|
+| `dd-trace-api/` | **Public** (published Maven artifact) | User-facing API: `@Trace`, `GlobalTracer`, `DDTags`, config constants, product APIs (LLMObs, AppSec, etc.) |
+| `internal-api/` | **Internal** (not published) | Shared internals: `AgentSpan`, `AgentTracer`, `Config`, `TagMap`, caches, utilities |
+| `dd-trace-core/` | **Internal** | Core engine: `DDSpan`, `CoreTracer`, serializers, writers |
+
+### Rules
+- Never add internal types to `dd-trace-api/` — it has zero dependencies on implementation.
+- Use `CharSequence` over `String` in internal APIs — enables `UTF8BytesString` passthrough without allocation.
+- Every public API must have a **NoOp implementation** for when the agent is not loaded or a feature is disabled.
+- **Deprecation**: annotate with `@Deprecated` + Javadoc pointing to the replacement + provide a default method that delegates to the new API. Never remove without a deprecation cycle.
+- Config keys use dot-separated lowercase without `dd.` prefix (e.g., `trace.agent.url`). The `DD_` env var prefix is applied automatically.