AI Engineer Paris 2026 Main Stage: Google DeepMind, ElevenLabs, Hugging Face & Stripe | Day 2
This conference recording surfaces three useful checks for agent builders: shorten integration paths, tune inference for the workload, and reject tests that merely restate implementation details.
The excerpts report **24% month-over-month** growth in iOS launches after agentic coding tools appeared and under six weeks from sandbox to first charge. An inference team reports **5.5×** more tokens per minute per GPU and about **400 vs 130 tokens/s** per user after workload tuning.
Build agent-native integrations that avoid dashboard handoffs, but measure the path to a real outcome rather than code produced. Tune serving separately for interactive agents and batch processing, and make generated tests exercise behavior instead of constants, source order, or mocked-away failures.
The excerpts report **24% month-over-month** growth in iOS launches after agentic coding tools appeared and under six weeks from sandbox to first charge. An inference team reports **5.5×** more tokens per minute per GPU and about **400 vs 130 tokens/s** per user after workload tuning. Build agent-native integrations that avoid dashboard handoffs, but measure the path to a real outcome rather than code produced. Tune serving separately for interactive agents and batch processing, and make generated tests exercise behavior instead of constants, source order, or mocked-away failures. This is a partial transcript spanning unrelated talks and sponsor presentations. The adoption and performance figures are presenter-reported, their methodologies are absent here, and the recording does not establish that the reported improvements generalize.
This shifts the coding-agent case from output volume to end-to-end outcomes: adoption speed matters only if work reaches a real charge or other completed result, and serving gains depend on workload-specific tuning. It also reinforces that generated tests are not automatically useful; they must exercise behavior and preserve real failure paths. The reported growth and throughput remain local claims without enough methodology to generalize.