At its inaugural Code with Claude conference on May 6, 2026, Anthropic released managed agents with sandboxed code execution, checkpointing, and credential scoping, alongside routines that trigger prompts on cron schedules, GitHub webhooks, or API endpoints.1 InfoQ 2026-05-07 Code with Claude on May 6, 2026 unveiled managed agents (sandboxed execution, checkpointing, credential scoping) and routines on cron/webhooks/APIs with classifier-based safety gating; Amodei cited 80x annualized Q1 growth and ~$30B annualized revenue. Open source Notably, it released no new model.2 MindStudio 2026-05-07 Five agent features including multi-agent orchestration, an Outcomes grading agent (8.4% Word, 10.1% slides improvement), Dreaming memory curation, Claude Finance's ten prebuilt agents, and add-ins; no new model released, infrastructure framed as the bottleneck. Open source 3 Simon Willison (live blog) 2026-05-06 Anthropic doubled Claude Code's five-hour limit for paid tiers, reported 17x year-over-year platform API volume, and described routines as higher-order prompts producing ready-to-merge pull requests. Open source The stake is a strategic reframing: Anthropic is arguing that the bottleneck for production AI coding is no longer raw model capability but the infrastructure to run agents safely and unattended.2 MindStudio 2026-05-07 Five agent features including multi-agent orchestration, an Outcomes grading agent (8.4% Word, 10.1% slides improvement), Dreaming memory curation, Claude Finance's ten prebuilt agents, and add-ins; no new model released, infrastructure framed as the bottleneck. Open source We assess with moderate confidence that this is a deliberate move up the stack, from selling a smart autocomplete to selling an operations platform, and that its success depends less on benchmarks than on whether teams trust agents to act without supervision.

The announcement was infrastructure, on purpose

The feature set clusters around running agents rather than making them smarter. Managed agents provide the sandbox, checkpoints, and scoped credentials that a system needs before it can execute code in production.1 InfoQ 2026-05-07 Code with Claude on May 6, 2026 unveiled managed agents (sandboxed execution, checkpointing, credential scoping) and routines on cron/webhooks/APIs with classifier-based safety gating; Amodei cited 80x annualized Q1 growth and ~$30B annualized revenue. Open source Routines let a developer schedule prompts or fire them from a GitHub event, and an auto mode uses classifiers to screen for destructive actions and prompt-injection risk before execution.1 InfoQ 2026-05-07 Code with Claude on May 6, 2026 unveiled managed agents (sandboxed execution, checkpointing, credential scoping) and routines on cron/webhooks/APIs with classifier-based safety gating; Amodei cited 80x annualized Q1 growth and ~$30B annualized revenue. Open source Multi-agent orchestration lets a lead agent delegate to specialist sub-agents working in parallel on a shared file system, with the activity auditable in the console.2 MindStudio 2026-05-07 Five agent features including multi-agent orchestration, an Outcomes grading agent (8.4% Word, 10.1% slides improvement), Dreaming memory curation, Claude Finance's ten prebuilt agents, and add-ins; no new model released, infrastructure framed as the bottleneck. Open source 3 Simon Willison (live blog) 2026-05-06 Anthropic doubled Claude Code's five-hour limit for paid tiers, reported 17x year-over-year platform API volume, and described routines as higher-order prompts producing ready-to-merge pull requests. Open source An Outcomes feature adds a separate grading agent that scores work against a rubric, which Anthropic said lifted quality by 8.4% on Word documents and 10.1% on slides.2 MindStudio 2026-05-07 Five agent features including multi-agent orchestration, an Outcomes grading agent (8.4% Word, 10.1% slides improvement), Dreaming memory curation, Claude Finance's ten prebuilt agents, and add-ins; no new model released, infrastructure framed as the bottleneck. Open source A research preview called Dreaming curates memory between sessions so agents refine over time.2 MindStudio 2026-05-07 Five agent features including multi-agent orchestration, an Outcomes grading agent (8.4% Word, 10.1% slides improvement), Dreaming memory curation, Claude Finance's ten prebuilt agents, and add-ins; no new model released, infrastructure framed as the bottleneck. Open source 3 Simon Willison (live blog) 2026-05-06 Anthropic doubled Claude Code's five-hour limit for paid tiers, reported 17x year-over-year platform API volume, and described routines as higher-order prompts producing ready-to-merge pull requests. Open source The through line, stated plainly by the company, is that infrastructure around existing models is the current constraint.2 MindStudio 2026-05-07 Five agent features including multi-agent orchestration, an Outcomes grading agent (8.4% Word, 10.1% slides improvement), Dreaming memory curation, Claude Finance's ten prebuilt agents, and add-ins; no new model released, infrastructure framed as the bottleneck. Open source

The business context explains the urgency. Anthropic said first-quarter 2026 usage and revenue grew 80x on an annualized basis, far above a planned 10x, with annualized revenue near $30 billion by early April, and platform API volume up 17x year over year.1 InfoQ 2026-05-07 Code with Claude on May 6, 2026 unveiled managed agents (sandboxed execution, checkpointing, credential scoping) and routines on cron/webhooks/APIs with classifier-based safety gating; Amodei cited 80x annualized Q1 growth and ~$30B annualized revenue. Open source 3 Simon Willison (live blog) 2026-05-06 Anthropic doubled Claude Code's five-hour limit for paid tiers, reported 17x year-over-year platform API volume, and described routines as higher-order prompts producing ready-to-merge pull requests. Open source Those figures are the company's own and describe a growth rate, not audited financials, so read them as directional.1 InfoQ 2026-05-07 Code with Claude on May 6, 2026 unveiled managed agents (sandboxed execution, checkpointing, credential scoping) and routines on cron/webhooks/APIs with classifier-based safety gating; Amodei cited 80x annualized Q1 growth and ~$30B annualized revenue. Open source Anthropic also doubled Claude Code's five-hour usage window for Pro, Max, and Enterprise customers, a concession that heavy agentic use was hitting limits.3 Simon Willison (live blog) 2026-05-06 Anthropic doubled Claude Code's five-hour limit for paid tiers, reported 17x year-over-year platform API volume, and described routines as higher-order prompts producing ready-to-merge pull requests. Open source

The target is Copilot's position, not its features

Anthropic is not trying to out-feature an autocomplete. It is trying to change what a developer buys. Claude Finance, a pack of ten prebuilt agents with connectors from Dun and Bradstreet, Fiscal AI, and Verisk, and add-ins that let Claude operate natively inside software like Word, point at selling vertical, deployable agents rather than a general assistant.2 MindStudio 2026-05-07 Five agent features including multi-agent orchestration, an Outcomes grading agent (8.4% Word, 10.1% slides improvement), Dreaming memory curation, Claude Finance's ten prebuilt agents, and add-ins; no new model released, infrastructure framed as the bottleneck. Open source Who gains from this is specific. Enterprises with backlogs of repetitive engineering and back-office work gain a way to run scheduled, audited agents against them.2 MindStudio 2026-05-07 Five agent features including multi-agent orchestration, an Outcomes grading agent (8.4% Word, 10.1% slides improvement), Dreaming memory curation, Claude Finance's ten prebuilt agents, and add-ins; no new model released, infrastructure framed as the bottleneck. Open source 1 InfoQ 2026-05-07 Code with Claude on May 6, 2026 unveiled managed agents (sandboxed execution, checkpointing, credential scoping) and routines on cron/webhooks/APIs with classifier-based safety gating; Amodei cited 80x annualized Q1 growth and ~$30B annualized revenue. Open source Anthropic gains usage-based revenue that scales with agent runtime rather than seat count, which its 80x growth figure suggests it is already capturing.1 InfoQ 2026-05-07 Code with Claude on May 6, 2026 unveiled managed agents (sandboxed execution, checkpointing, credential scoping) and routines on cron/webhooks/APIs with classifier-based safety gating; Amodei cited 80x annualized Q1 growth and ~$30B annualized revenue. Open source

Who is pressured is equally specific. GitHub Copilot's seat-based model is the incumbent that a runtime-billed agent platform undercuts, though the picture is complicated because GitHub already ships Claude to its own developers through chat, CLI, and code review, making Anthropic both supplier and competitor.3 Simon Willison (live blog) 2026-05-06 Anthropic doubled Claude Code's five-hour limit for paid tiers, reported 17x year-over-year platform API volume, and described routines as higher-order prompts producing ready-to-merge pull requests. Open source Pure code-review and CI startups are squeezed by Anthropic folding review, autofix, and security scanning into the platform.2 MindStudio 2026-05-07 Five agent features including multi-agent orchestration, an Outcomes grading agent (8.4% Word, 10.1% slides improvement), Dreaming memory curation, Claude Finance's ten prebuilt agents, and add-ins; no new model released, infrastructure framed as the bottleneck. Open source 3 Simon Willison (live blog) 2026-05-06 Anthropic doubled Claude Code's five-hour limit for paid tiers, reported 17x year-over-year platform API volume, and described routines as higher-order prompts producing ready-to-merge pull requests. Open source The party carrying the new risk is the customer: unattended agents with scoped credentials acting on production systems concentrate operational and security exposure, which is exactly why the classifier-based safety gating shipped alongside the automation.1 InfoQ 2026-05-07 Code with Claude on May 6, 2026 unveiled managed agents (sandboxed execution, checkpointing, credential scoping) and routines on cron/webhooks/APIs with classifier-based safety gating; Amodei cited 80x annualized Q1 growth and ~$30B annualized revenue. Open source

The counter-case

The platform thesis could be wrong if trust, not tooling, is the real barrier. Sandboxes and prompt-injection classifiers reduce risk but do not eliminate it, and a single high-profile incident of an autonomous agent taking a destructive action on production could set unattended adoption back regardless of how good the infrastructure is.1 InfoQ 2026-05-07 Code with Claude on May 6, 2026 unveiled managed agents (sandboxed execution, checkpointing, credential scoping) and routines on cron/webhooks/APIs with classifier-based safety gating; Amodei cited 80x annualized Q1 growth and ~$30B annualized revenue. Open source The absence of a new model is also double-edged: if a rival ships a materially stronger coding model, Anthropic's infrastructure advantage may not hold buyers who chase capability.2 MindStudio 2026-05-07 Five agent features including multi-agent orchestration, an Outcomes grading agent (8.4% Word, 10.1% slides improvement), Dreaming memory curation, Claude Finance's ten prebuilt agents, and add-ins; no new model released, infrastructure framed as the bottleneck. Open source And the growth figures that justify the strategy are self-reported rates over a small base, which can decelerate sharply.1 InfoQ 2026-05-07 Code with Claude on May 6, 2026 unveiled managed agents (sandboxed execution, checkpointing, credential scoping) and routines on cron/webhooks/APIs with classifier-based safety gating; Amodei cited 80x annualized Q1 growth and ~$30B annualized revenue. Open source For the thesis to fail, enterprises would have to keep agents on a human leash, buying the tooling but not granting the autonomy that makes the platform economics work.

What to watch

  • Unattended adoption is disclosed. Watch for named enterprises reporting production workloads run by managed agents without a human approving each action within two quarters; that, not feature counts, is the real test of the platform bet.1 InfoQ 2026-05-07 Code with Claude on May 6, 2026 unveiled managed agents (sandboxed execution, checkpointing, credential scoping) and routines on cron/webhooks/APIs with classifier-based safety gating; Amodei cited 80x annualized Q1 growth and ~$30B annualized revenue. Open source 2 MindStudio 2026-05-07 Five agent features including multi-agent orchestration, an Outcomes grading agent (8.4% Word, 10.1% slides improvement), Dreaming memory curation, Claude Finance's ten prebuilt agents, and add-ins; no new model released, infrastructure framed as the bottleneck. Open source
  • The GitHub relationship strains. Track whether Anthropic's direct agent platform and GitHub's Copilot business openly compete by year end, or whether GitHub deepens its reliance on Claude; the direction reveals who controls the developer surface.3 Simon Willison (live blog) 2026-05-06 Anthropic doubled Claude Code's five-hour limit for paid tiers, reported 17x year-over-year platform API volume, and described routines as higher-order prompts producing ready-to-merge pull requests. Open source
  • A safety incident, or its absence. If a public failure of an autonomous coding agent surfaces over the next two to three quarters, expect the industry to tighten autonomy defaults; a clean run would accelerate the shift Anthropic is selling.1 InfoQ 2026-05-07 Code with Claude on May 6, 2026 unveiled managed agents (sandboxed execution, checkpointing, credential scoping) and routines on cron/webhooks/APIs with classifier-based safety gating; Amodei cited 80x annualized Q1 growth and ~$30B annualized revenue. Open source
  • Growth rate holds. If Anthropic's next disclosure still shows usage growth well above its 10x plan, the runtime-billed model is compounding; a sharp deceleration would suggest the 80x figure was an early-base artifact.1 InfoQ 2026-05-07 Code with Claude on May 6, 2026 unveiled managed agents (sandboxed execution, checkpointing, credential scoping) and routines on cron/webhooks/APIs with classifier-based safety gating; Amodei cited 80x annualized Q1 growth and ~$30B annualized revenue. Open source 3 Simon Willison (live blog) 2026-05-06 Anthropic doubled Claude Code's five-hour limit for paid tiers, reported 17x year-over-year platform API volume, and described routines as higher-order prompts producing ready-to-merge pull requests. Open source

The forward implication is that the coding-tools contest is moving from whose model writes better code to whose infrastructure can be trusted to run it. Anthropic has placed its bet on the second question, and the market will answer it not with benchmark scores but with how much authority customers are willing to hand an agent.