Anthropic's Code with Claude developer conference in San Francisco on May 6, 2026 dropped a wave of platform features aimed squarely at coding teams, and the hosts walk through what actually matters for builders. Devyatkin reads the SpaceX–Anthropic compute deal as one of the cleverest business moves of the year — xAI is sitting on underutilized GPUs, and "the enemy of my enemy" gets to raise everyone's usage limits. The episode dissects routines, outcomes/goals, multi-agent orchestration, the advisor pattern, dreaming, and the new Anthropic-on-AWS path that is *not* Bedrock, with Gonçalves flagging an easy-to-miss compliance gotcha: on that path, data still leaves AWS for Anthropic. The closing read on OpenAI is sharp — a two-month free Codex trial for new Codex users on eligible enterprise accounts, which the hosts call "a little bit of a desperate attempt" after Anthropic's enterprise adoption crossed OpenAI's.
Summary
Anthropic’s Code with Claude developer conference in San Francisco on May 6, 2026 dropped a wave of platform features aimed squarely at coding teams, and the hosts walk through what actually matters for builders. Devyatkin reads the SpaceX–Anthropic compute deal as one of the cleverest business moves of the year — xAI is sitting on underutilized GPUs, and “the enemy of my enemy” gets to raise everyone’s usage limits. The episode dissects routines, outcomes/goals, multi-agent orchestration, the advisor pattern, dreaming, and the new Anthropic-on-AWS path that is not Bedrock, with Gonçalves flagging an easy-to-miss compliance gotcha: on that path, data still leaves AWS for Anthropic. The closing read on OpenAI is sharp — a two-month free Codex trial for new Codex users on eligible enterprise accounts, which the hosts call “a little bit of a desperate attempt” after Anthropic’s enterprise adoption crossed OpenAI’s.
Key Topics
The SpaceX compute deal and the race for GPUs
The episode opens on a piece of news adjacent to the conference itself: SpaceX is giving Anthropic access to one of its data centers, opening up enough compute that Anthropic could raise per-user usage limits across the board. Devyatkin’s read is that the deal is unusually clever from a business standpoint. xAI built large GPU capacity for training Grok, but Grok’s traction has not kept pace, leaving racks of expensive hardware underutilized. Samoylov adds the operational angle — those video cards “can’t just shut down,” they have to be powered and ready for load — so monetizing them by leasing capacity to a rival is the rational move.
The hosts also note reporting that Cursor is in conversations with SpaceX around investment and potential acquisition, with SpaceX-hosted compute as part of the arrangement — effectively positioning SpaceX as a new compute platform on the eve of going public. The DevOps tangent writes itself: if data centers eventually move to orbit, deploying software gets a new failure mode. As Samoylov puts it, “in space you get bigger luck.” Treat the space-data-center riff as a thought experiment rather than a near-term plan — and a useful illustration of how far ahead infrastructure-as-code thinking has to look.
Routines: cron jobs for Claude Code
Routines are scheduled Claude Code runs — effectively cron jobs for an agent. They can run locally (laptop awake, terminal up) or shift to the cloud. Devyatkin describes a concrete workflow he already runs: a weekly skill that crawls every B.O.R.I.S repository, summarizes development activity, updates the knowledge base, and posts to Slack for the marketing side. He schedules it during meeting blocks where the laptop is guaranteed to be alive, and a pull request lands by the end of the call.
Samoylov shares another routine: a daily dead-code check across repositories. With LLM-generated code accumulating fast, files get orphaned quickly, and a scheduled sweep catches what manual review misses. Devyatkin extends the pattern to multi-repo CI: GitHub Actions has notoriously poor cross-repo visibility (“the only thing I miss from Jenkins is the unified dashboard”), and a scheduled agent that pulls all failing workflows into a single report fills that gap without rolling your own automation.
The framing the hosts return to: Anthropic is watching what users built around the harness — OpenClaw and Hermes Agent are referenced (covered in episode #7) — and absorbing the popular patterns into the platform so fewer teams have to build them from scratch.
Outcomes and goals: long-running work on the platform
Outcomes is a new platform-level capability where a developer hands Anthropic a goal and the platform iterates on it server-side until it is done or needs input. Inside Claude Code the same idea surfaces as “goals.” Gonçalves connects it to “super plan,” which uses platform-side multi-agent work for plans too large for a single session.
Devyatkin’s mental model: this is the Ralph Wiggum loop pattern — “make the LLM go until it’s done” — productized. Teams used to wire that loop with their own hooks; now the platform owns it. The practical upside for builders is that the loop, the multi-agent split, and the state management run on Anthropic’s infrastructure instead of pinning a local process for hours.
Multi-agent orchestration via the task tool
The orchestration model: an orchestrator agent splits a task across sub-agents, each with its own context window. Initial artifacts are shared, but per-agent context isolation keeps the orchestrator from drowning in detail. Gonçalves walks through a practical example — documenting multiple projects in a single session. With no instruction, Claude Code defaulted to one serial session. Once he explicitly said “launch four agents, one per project, produce documentation in this format,” the four sub-agents finished in parallel in three minutes, and the orchestrator stitched cross-project links back together once they returned.
The takeaway is operational, not architectural: if work decomposes cleanly, tell the agent to use sub-agents. The harness does not always infer parallelism on its own.
Anthropic platform on AWS (and why it is not Bedrock)
Anthropic announced availability of its platform on AWS, distinct from Bedrock. Devyatkin’s read — caveated as a take to verify — is that this is effectively AWS acting as a proxy in front of the Anthropic platform: requests originate inside an AWS account, use IAM and AWS billing, but the underlying calls route to Anthropic. The procurement and identity benefits are concrete — no separate Anthropic credentials to provision, billing rolls up under AWS, and Claude Managed Agents is available on the path. Feature parity, though, is not complete: per Anthropic’s docs, outcomes, multi-agent sessions, and webhooks are not currently available on Claude Platform on AWS. Builders designing around those capabilities should plan to call the Anthropic API directly until that gap closes.
Samoylov flags an important data-residency distinction the hosts want listeners to internalize: on Bedrock, data stays within AWS and is not shared with the model provider; on the new Anthropic-on-AWS path, data is still sent to Anthropic. Same models, different compliance posture. If a regulated workload was approved for Bedrock specifically because of that boundary, the new path is not a drop-in equivalent.
This came up first in episode #7; the hosts treat it as worth re-stating because the naming invites misreading.
The advisor pattern: Opus advising Sonnet
Anthropic introduced an advisor pattern: a more capable model — Opus — can advise a mid-tier model — Sonnet — on harder steps, so a workflow does not have to pay Opus prices for every token. With B.O.R.I.S, Gonçalves notes the multi-model approach has been there from day one: “you are not going to throw the most expensive model at everything.” The advisor pattern formalizes that hierarchy at the API.
Devyatkin offers a caveat for Bedrock users about where the advisor tool actually runs. Per Anthropic’s documentation, the advisor tool is available on the Claude API and on Claude Platform on AWS, but is not available on Amazon Bedrock, Vertex AI, or Microsoft Foundry. The broader pattern the hosts have seen with new platform capabilities is that they ship via request headers (extended thinking landed the same way), and provider invoke surfaces like Bedrock’s do not always expose those headers on day one. Builders who need the advisor pattern today should target the Claude API or Claude Platform on AWS, and treat Bedrock, Vertex AI, and Microsoft Foundry as separate compatibility questions — worth verifying against the docs for the specific feature before designing around it.
The broader observation Samoylov makes: AI gateways and proxy products gain “use any model” simplicity but lose feature parity, because feature flags arrive via provider-specific headers that a generic gateway does not pass through.
Webhooks for long-running outcomes
With outcomes running server-side and potentially taking a long time, polling is wasteful. Anthropic announced webhooks so the platform can notify your application when an outcome completes or stalls waiting for input. Builders using the SDK get the same event delivery without rolling their own queue.
Dreaming: memory consolidation on the platform
“Dreaming” is the platform-side memory consolidation feature — what Claude Code already does locally when it compacts a session into long-term notes is now available through the API. The metaphor is deliberate: human brains prune during sleep, keeping what is useful and discarding what is not. Devyatkin reaches for the Sherlock Holmes line about not bothering to learn the Earth orbits the Sun if it does not help him solve cases — an agent’s memory should serve the job, not accumulate trivia.
This is the same lifecycle the hosts laid out in episode #7: capture, manage, retrieve. The conference announcement moves “manage” — the hardest of the three to implement well — onto the platform.
Agent view and the convergence with other harnesses
Claude Code also gained an agent view: running, completed, and waiting-for-input agents in a single pane, similar to what Cursor surfaces. Gonçalves frames this as part of the broader convergence — different harnesses are landing on the same UI primitives because the underlying workflows are the same.
The OpenAI counter-move
The episode closes on a competitive reading. Anthropic’s enterprise adoption recently crossed OpenAI’s on a widely-circulated chart, and OpenAI’s response includes a two-month free Codex offer scoped to new Codex users on eligible enterprise accounts — so organizations that are already OpenAI customers can still qualify, but only for users new to Codex. Devyatkin’s read is candid: “it does look a little bit desperate.” Samoylov adds the telecom analogy — existing paying Codex customers get nothing while the promotion goes to new signups on the product. Both hosts land on the same upside: competition keeps everyone honest. If only one provider held the frontier, “privacy will be skyrocketing.” The presence of Chinese open-weight models gets a brief nod for keeping the closed providers in line on pricing.
Resources
- Code with Claude — San Francisco, May 6, 2026 — Official event page for the San Francisco stop of Anthropic’s 2026 developer conference, the source of the announcements covered here; London (May 19) and Tokyo editions follow.
- Claude Code documentation — Official docs for the harness, including routines, goals, sub-agents, and the file-based memory system referenced throughout.
- Claude Managed Agents overview — The managed-agents docs that cover outcomes, multi-agent orchestration, dreaming, and webhooks, and where availability on Claude Platform on AWS is documented.
- Anthropic models on Bedrock — Background on running Anthropic models inside AWS Bedrock, where data stays within AWS; useful as the contrast point for the new Anthropic-platform-on-AWS path the hosts discuss.
- GitHub CLI (
gh) — The tool that makes scheduled cross-repo CI inspection workable from a Claude Code routine. - Episode #4 — Harness Engineering — Background on the harness/platform split that explains why outcomes, goals, and dreaming surface differently in Claude Code than in the API.
- Episode #6 — The Big AI Squeeze — Context for the OpenAI Codex pricing move and the compute economics behind the SpaceX deal.
- Episode #7 — When Agent Memory Helps and When It Hurts — The memory lifecycle the platform’s new “dreaming” feature operationalizes, including the Bedrock-vs-Anthropic-platform compliance distinction the hosts return to here.
Sources: