Sakana AI launched Fugu today, a production multi-agent orchestration system that dynamically routes tasks across a swappable pool of frontier models while presenting a single API to the caller.

• Benchmark parity without frontier training: Sakana claims Fugu matches or outperforms Anthropic's Fable and Mythos on select benchmarks by coordinating existing models rather than training a new one — a direct architectural bet against the scale-is-everything thesis.
• Export-control hedge baked in: The system's swappable model pool means operators can replace any model that goes offline due to regulatory action, a design choice that resonates given the recent Fable 5 and Mythos 5 suspension saga.

Zilliz launched public preview of Vector Lakebase on June 22, extending Milvus beyond vector search into a unified data platform that covers ingestion, storage, and retrieval for AI pipelines.

• Fragmentation is the stated target: The product attacks the multi-tool sprawl in enterprise AI data stacks — separate pipelines for embedding, chunking, storing, and retrieving have been the dominant pain point in production RAG deployments.
• Milvus installed base as distribution lever: With Milvus cited as the most widely adopted open-source vector database, Zilliz is using the existing developer footprint to pull teams toward a managed, consolidated offering rather than starting from scratch.

Nylas shipped general availability of Agent Accounts on June 22, giving AI agents a provisioned email address and calendar identity managed entirely through a single API call.

• Closes the action gap in agentic workflows: Agents that can reason and plan have lacked a credible communications identity to act from — this makes agent-sent emails first-class messages rather than spoofed or shared addresses that create audit and deliverability problems.
• Enterprise governance implication: A discrete, API-managed identity per agent means access logs, revocation, and compliance trails attach to the agent rather than bleeding into a human employee's mailbox.

GitHub updated its Copilot usage metrics REST API on June 19 with an ai_credits_used field that exposes per-developer, per-day AI credit consumption under the usage-based billing model that launched June 1.

• Spend visibility was the missing control: GitHub switched all Copilot plans from flat-fee to token-metered billing on June 1, but org admins had no per-user view of who was drawing down the credit pool until this update.
• Agentic usage is the driver: The billing shift was explicitly triggered by agentic runs, which can consume orders of magnitude more tokens than a line completion — the new field is the first signal that GitHub expects finance and IT to treat Copilot like a cloud service with showback.

Adobe reported Q2 revenue of $6.62 billion, up 13% year-over-year, with AI-first ARR tripling to surpass $500M and Firefly alone reaching $300M ARR growing roughly 50% quarter-over-quarter.

• Margins expanding through AI absorption: Non-GAAP EPS grew 18% to $5.96 and net margins sit near an all-time high of 36%, meaning Adobe is paying for generative AI compute while still widening profitability — the opposite of the cost-curve fear that drove its stock lower.
• Freemium funnel validating the strategy: Freemium MAUs grew from 700M to 850M, suggesting AI features are pulling new users into the Adobe ecosystem who then convert to paid, rather than cannibalizing existing subscribers.