self.md radar — 2026-04-29
The assistant is slipping out of the vendor-owned app and into shared control planes: AWS wants OpenAI inside Bedrock, GitHub is rebuilding the repo host around agent traffic, and Brussels wants Android’s assistant slot opened beyond Gemini.
Today leads with AWS turning OpenAI into Bedrock inventory plus managed agents and Codex inside the same procurement stack. Then GitHub admits agentic development is already forcing a 30X-scale redesign of the repo host itself. Closing on the EU treating Android AI integration as an interoperability surface rather than a Google-only privilege.
1. AWS puts OpenAI models, Codex, and managed agents on Bedrock
sources:
what happened: Amazon announced three limited-preview offerings in a single move: OpenAI models on Amazon Bedrock, Codex on Bedrock, and Bedrock Managed Agents powered by the OpenAI harness. AWS says customers can now reach OpenAI frontier models through the same Bedrock services they already use for access, fine-tuning, and orchestration alongside Anthropic, Meta, Mistral, Cohere, and Amazon’s own models. GPT-5.5 and GPT-5.4 are slated for preview on Bedrock with unified security, governance, and cost controls, and the Managed Agents runtime keeps inference, memory, and skills inside the customer environment with data that never leaves AWS.
why this matters: Hyperscaler competition is shifting from exclusive model access to who owns the enterprise control plane around the models, and AWS just claimed OpenAI as inventory inside its own procurement, governance, and agent runtime stack.
2. GitHub is rebuilding for 30X scale because of agent traffic
sources:
what happened: GitHub says it started a 10X capacity plan in October 2025, but by February 2026 it was clear the platform needed to design for a future requiring 30X today’s scale. The company explicitly names agentic development workflows accelerating sharply since the second half of December 2025 as the main driver, with repository creation, pull request activity, API usage, automation, and large-repository workloads all climbing fast. GitHub says this is forcing changes across caching, service isolation, a monolith-to-Go migration, webhook backends, authentication flows, additional compute on Azure, and a path toward multi-cloud.
why this matters: AI traffic is now reshaping the core architecture of the dominant code host, which means the substrate every shipping team depends on is being rebuilt around agent throughput rather than human commit cadence.
3. The EU wants Android’s AI surface opened to non-Google assistants
sources:
what happened: The European Commission issued preliminary findings and draft measures under the DMA focused on Android interoperability for competing AI services. The Commission says competing assistants should be able to interact with Android apps and execute tasks such as sending email, ordering food, or sharing photos, and the draft measures would also let users activate them with a custom wake word. Comments are open until 2026-05-13, with a final binding decision due within six months of the specification proceedings opening; Ars reports Google calling the move “unwarranted intervention” that would mandate access to sensitive hardware and device permissions.
why this matters: The assistant slot on a phone is being treated like a regulated interoperability surface rather than a private default, which puts the most valuable consumer agent entrypoint under the same kind of access mandate that once reshaped browsers and search.
supporting links
- Long-running Agents — persistence, recovery, and session handoff are now first-class harness problems, not glue around a chat loop.
- Mistral Workflows — public-preview orchestration layer with durability, observability, and human approval built into the substrate.
- Temporal — neutral durable-execution layer that AI workflow vendors keep landing on top of.
- trycua/cua — background computer-use for macOS, the agent surface pushing into native app control rather than a chat panel.
left on the table
- talkie — same provenance and public-domain model story as yesterday’s edition with no hard new delta today.
- OpenCLI — same “turn arbitrary surfaces into agent interfaces” wave as the Apr 17 Kampala signal, weaker and more repo-thin.
- Glance Dashboard V2 GA — useful self-hosted artifact but the same observability and dashboard lane already led Apr 19.
- Qwen 3.6 27B quant eval — real operator benchmark but the local-model evaluation lane was covered Apr 21 and the Qwen release lane Apr 23.
- A good AGENTS.md is a model upgrade — interesting evidence that repo-local instructions shape model performance, but too close to the recent harness and context-control wave to promote.