adk-python — Engineering Performance
19 engineers all time · Apr 2025 – Nov 2026 · built 2026-11-22 · GitHub
Performance snapshot
Today's rolling 90-day reading for adk-python, compared with the start of the series. Pick a window to move that comparison point.
Avg. perf / dev / mo
+285.6%
1.06 → 4.09 ETV
Active engineers
+40.0%
10.0 → 14.0
Features
−7.8pp
20.7% → 12.9%
vs. Google
1.1x
1.2x → 1.1x · +11% above
adk-python vs. Google
Per-engineer ETV for adk-python against Google as a whole. Both lines are 90-day rolling averages scaled to a 30-day month, so they share one axis and can be read against each other at any point. Pick a window to zoom the chart to it.
Performance Composition
Each month's output split by type of work: Features (new value), Maintenance (sustaining systems), Tests, Docs, and Fixes (rework). The yellow line is output per engineer, so when it rises each engineer is delivering more, whatever the team size did. Unit: Engineering Throughput Value (ETV).
Engineering capacity
Effective engineers behind adk-python, against its pre-AI baseline. Each subject has its own: adk-python's is 1.06 ETV / dev / mo, its first reading in Q2 2025. Per-engineer ETV divided by that gives a capacity multiple, and that multiple applied to the engineers active in the trailing 90 days turns it into engineer-equivalents. The line is the real headcount, so the gap between line and area is what the leverage is worth. Because each baseline is its own, every subject opens at 1.0x on its first day: multiples measure improvement and are not comparable between subjects.
Knowledge concentration
How dependent is this repo on a small number of engineers? Higher top-1 share = higher key-person risk.
George Weale owns 32.6 % of commits.
Behind the numbers
Written summary of the work completed each month.
No monthly reports available yet.
Top engineers
Most impactful commits
Top 10 by ETV in the all-time window.
- 12.9ETVtest: add unit tests for public symbols that had no coverage Co-authored-by: George Weale <gweale@google.com> PiperOrigin-RevId: 960421043George Weale · 456524d7 · 2026-08-06
- 10.8ETVrefactor(types): make the agents, sessions, flows and workflow packages pass strict mypy Not annotations-only. This is one component's slice of a repo-wide typing cleanup, and the wider change was found to contain behavior changes that have not all been individually triaged, so please review it as a functional change. Co-authored-by: George Weale <gweale@google.com> PiperOrigin-RevId: 958554313George Weale · 07add3b8 · 2026-08-03
- 3.6ETVfeat(sessions): add BufferedSessionService to write session events in the background Co-authored-by: Ayush Agrawal <ayushagra@google.com> PiperOrigin-RevId: 1023012449Ayush Agrawal · e1deb17f · 2026-11-16
- 3.2ETVrefactor(types): make google.adk.models pass strict mypy Not annotations-only. This is one component's slice of a repo-wide typing cleanup, and the wider change was found to contain behavior changes that have not all been individually triaged, so please review it as a functional change. Co-authored-by: George Weale <gweale@google.com> PiperOrigin-RevId: 958623646George Weale · 7f82142a · 2026-08-03
- 2.8ETVrefactor(types): type the integrations, skills and MCP tool packages for strict mypy Co-authored-by: George Weale <gweale@google.com> PiperOrigin-RevId: 961125207George Weale · cd36dbc3 · 2026-08-07
- 2.7ETVfeat(live): add Gemini speech-to-text and text-to-speech for CascadeLive Adds `GeminiSTT` and `GeminiTTS`, the `LiveIngress` and `LiveEgress` transforms backed by Gemini, so a `CascadeLive` agent can run recognition, reasoning and synthesis on Gemini with one set of credentials and no extra dependency. Both are built for low latency. - `GeminiSTT` streams the user's audio to `gemini-3.5-transcribe-live` over the Live API, so speech is transcribed while the user talks. The server's voice activity detection reports `UserSpeechStarted`, interim transcripts become `PartialTranscript` captions, and the final transcript becomes `UserTurnFinished` after 500 ms of silence by default. `UserSpeechStarted` lets `CascadeLive` interrupt the agent as soon as the user talks over it. If the client signals the end of the user's turn itself, for example by sending `ActivityEnd` when a push-to-talk button is released, the final transcript arrives right away instead of after the silence. Sessions are replaced before the Live API's 10-minute limit, between utterances or right away on GoAway, and a session that drops is reopened; either way the new session is sent the speech the old one had not finished transcribing, up to `max_replay_ms`, and the start of that speech is not reported twice. - `GeminiTTS` splits the reasoner's text into segments and synthesizes each with a streaming `generate_content` call to `gemini-3.8-flash-lite-tts`. The next segment is synthesized while the current one is delivered, so there are no gaps between segments, and the first reply warms up the connection while the reasoner thinks. A request that fails with a transient error before returning any audio is retried. The first segment ends at the first sentence end; later ones group sentences until they reach `min_segment_chars` (50), because a spelled-out letter such as `Y.` synthesized on its own is often not read as its name. - Both are experimental, behind the `CASCADE_LIVE` feature. - The `live/cascade` sample picks its speech transforms with the `CASCADE_SPEECH_PROVIDER` environment variable, `gemini` by default or `elevenlabs`, with a companion README for each provider. - Adds unit guides for `GeminiSTT` and `GeminiTTS`, referenced from the CascadeLive guide. Co-authored-by: Liang Wu <wuliang@google.com> PiperOrigin-RevId: 1000461641Liang Wu · f78149a9 · 2026-10-15
- 2.3ETVfeat: Migrate ADK to google-cloud-aiplatform v2.2 (agentplatform) Moves the google-cloud-aiplatform pin from >=1.148.1,<2 to >=2.2,<3 and migrates call sites to the v2 `agentplatform` surface (agent_engines -> runtimes; sessions, sandboxes and memory_banks move to the client; AdkApp -> agentplatform.frameworks). The floor is 2.2, not 2.1: 2.2 makes `vertexai.types` and `agentplatform.types` the same classes, so retrieve_profiles() keeps its public `list[vertex_types.MemoryProfile]` annotation. VertexAiSessionService and VertexAiMemoryBankService fall back to the legacy `agent_engines` path when a subclass's _get_api_client returns a `vertexai` client, which in 2.x has only that path; both paths take the same arguments and return the same types. Deploy CLI: AdkApp now reads project and region from the environment, so fast_api.py sets GOOGLE_CLOUD_PROJECT and GOOGLE_CLOUD_AGENT_ENGINE_LOCATION, and in express mode clears them. Deploy CLI: _ensure_agent_engine_dependency appends a >=2.2,<3 floor for each Agent Platform distribution an agent pins, and pip fails the image build if a pin conflicts with its floor. A hash-locked requirements file is left as written, since pip rejects unhashed requirements in that mode. _AGENT_ENGINE_CLASS_METHODS adds the 7 async artifact methods that v2 registers. VertexAiCodeExecutor stays on the legacy `vertexai` surface, which 2.x still ships, because agentplatform has no Extension equivalent. PiperOrigin-RevId: 995018206Amy Wu · 77d4dcd0 · 2026-10-07
- 2.1ETVfeat: Add ComputerUseToolset PiperOrigin-RevId: 785646071Xiang (Sean) Zhou · 083dcb44 · 2025-07-22
- 2.0ETVrefactor(types): make google.adk.a2a pass strict mypy Not annotations-only. This is one component's slice of a repo-wide typing cleanup, and the wider change was found to contain behavior changes that have not all been individually triaged, so please review it as a functional change. Co-authored-by: George Weale <gweale@google.com> PiperOrigin-RevId: 959932563George Weale · b34c636f · 2026-08-05
- 2.0ETVfeat: support context caching 1. add a context cache config in app level which will apply to all agents in the app 2. pass on cache config through invocation context to llm_reqeust 3. store cache metadata in llm_response 4. lookup old cache metadata from latest event for reusing old cache 5. create new cache if old cache cannot be reused PiperOrigin-RevId: 809158578Xiang (Sean) Zhou · c66245a3 · 2025-09-19