Anthropic traced recent Claude Code quality reports to three changes: default reasoning effort moved from high to medium, idle-session reasoning was repeatedly cleared by a bug, and a reduced-verbosity prompt hurt coding quality. The issues were fixed April 20.

Anthropic restored higher effort defaults because users prefer default intelligence and opt down for simpler tasks. Test-time compute is not just a model parameter; it is trust and product defaults.

Claude says it will standardize access at 50% usage for plans that use Fable most intensively and continue capacity investment. Users need predictable plan contracts, not vague unlimited promises.

OpenAI and Anthropic are dealing with the same problem: once top models enter daily work, usage limits, default effort, quality reports, and capacity updates become core product surfaces.

Claude Managed Agents can execute tools in sandboxes controlled by the enterprise. Code, sensitive files, packages, services, and data stay inside the perimeter while Anthropic runs the loop, orchestration, context, and recovery.

MCP tunnels connect Claude Managed Agents to private MCP servers: internal databases, APIs, knowledge bases, and ticketing systems become tools through an outbound gateway without inbound firewall rules.

Anthropic names Cloudflare, Daytona, Modal, and Vercel as sandbox providers. Agent execution is becoming its own ecosystem around isolation, startup, state, CPU/GPU, VPC, and credential injection.

Rauch posted sandbox downloads data and said it is time to ship more agents. Platform competition has expanded from model APIs to where agents safely run.

Peter Yang says staring at screens to manage agents all day fries the brain; better to walk outside, talk to agents on the phone, assign work, and get voice status updates. He is waiting for a lab to ship it.

Peter Steinberger asks whether we are still talking loops or have shifted to graphs. Agent orchestration is moving from simple loops and batches toward dependency graphs and parallel state.

Steipete watched Codex use browser and computer use to open Chrome, navigate a PR, tap a comment, and use the macOS picker just to upload an image. Lack of API does not stop agents, but it makes paths both amazing and painful.

Zara says people were uncomfortable recording meetings a few years ago; now business meetings are assumed recorded, not for humans but for agents. Agent adoption is changing cultural defaults.

Sachin describes AI data centers as self-contained infrastructure that primarily produces intelligence. Misunderstandings around water, power, communities, and jobs need engineering detail and local negotiation.

The episode explains OpenAI data centers are liquid cooled with closed-loop recycled water, small relative to household consumption. Infrastructure debates need technical facts, not vague impressions.

Site selection depends on land, permitting, grid/gas, construction labor, and electricians. AI speed is constrained by the physical world; model cadence must pass through these bottlenecks.

OpenAI acts as an offtake tenant across hyperscalers and neoclouds, buying guaranteed compute. Guaranteed capacity becomes the unit of intelligence supply, tying into capability, price, access, and limits.

As GPT-5.6 Sol and Claude Fable quality signals strengthen, users move real work in, pulling capacity, limits, default effort, and plan design into product discussions.

OpenAI’s compute episode shows physical buildout never feels fast enough; Claude Fable and Codex resets show access policy is the user interface of infrastructure reality.

Self-hosted sandboxes and MCP tunnels show enterprises will not simply hand data to cloud agents; execution, network, audit, credentials, and tools must stay governable.

Today’s feed points beyond magic prompts: high-quality agent use needs evals, harnesses, sandboxes, prototypes, voice control, AEO automation, and capacity planning.