Performance¶

How fast the SDK's hot paths are, measured with scripts/benchmark.php. It has no extra dependencies, so you can run it on your own server:

php scripts/benchmark.php            # 2 s per case
php scripts/benchmark.php --seconds=5

The script refuses to time anything if the benchmark agent doesn't complete its task, so the numbers are never an error path.

Results¶

Measured 2026-09-26 on an AWS Lightsail instance: Intel Xeon Platinum 8259CL @ 2.50 GHz, 4 vCPU, 15 GiB RAM, PHP 8.4.25 (CLI, no JIT), pure-PHP protobuf runtime (google/protobuf, no ext-protobuf).

Case	Operations per second	Time per operation
ProtoJSON encode a `Task` (`serializeToJsonString`)	931	1,074 µs
ProtoJSON decode a `Task` (`mergeFromJsonString`)	2,876	348 µs
SSE parse, 4 events (`EventStreamParser`)	169,024	5.9 µs
JSON-RPC `SendMessage`, full request (in-memory stores)	834	1,199 µs
JSON-RPC `SendStreamingMessage`, full SSE body	663	1,509 µs

What each case does:

ProtoJSON encode / decode: a completed Task with 10 history messages and one artifact with 3 parts (3.8 KB of JSON).
SSE parse: the client's stream parser reading a real 4-event stream (1.1 KB) from the server below.
JSON-RPC SendMessage: the whole server path in one process. A PSR-7 request goes through JsonRpcDispatcher and DefaultRequestHandler to a small agent that publishes a Task, a status update and an artifact; in-memory stores; a JSON response. No network.
SendStreamingMessage: the same, producing the full SSE body.

What the numbers mean¶

The SDK adds about 1–1.5 ms per request in-process. Your agent's own work (an LLM call is typically 500 ms or more) and the network will dominate.
Protobuf JSON is the biggest cost. The pure-PHP runtime works everywhere. For heavy traffic, install the C extension (pecl install protobuf); the SDK uses it automatically, with no code changes.
Stores and queues change the picture. These numbers use in-memory stores. With PdoTaskStore / PdoQueueManager or Redis, each request adds a few database round trips.

ext-protobuf¶

ext-protobuf wasn't measured on the machine above: building it needs the PHP development headers, which are a system package. The manual Benchmark workflow runs both runtimes side by side on a GitHub runner, and the results appear in that run's summary.