| Hey there, I once lost a $200 rebate because the check sat in a drawer past its 90-day window. Free money, defaulted to zero. Anthropic's Agent SDK credit works the same way — except it expires into an outage, on Monday. Below: the two clicks that prevent it, plus Hermes and OpenClaw shipping the boring stuff that makes open-source agents deployable at work.
🔥 The Big Thing Following up: 3 days to the Claude billing cliff — the credit isn't automatic We led June 4 with Anthropic splitting agent traffic off your Claude subscription. This is the follow-up — and the last issue before it bites. The new detail changes your to-do list: the credit isn't automatic. What flips Monday. On June 15, the Agent SDK, claude -p scripts, Claude Code GitHub Actions, and third-party Agent SDK apps stop drawing on subscription usage. They move to a separate monthly Agent SDK credit pool: $20 on Pro, $100 on Max 5x, $200 on Max 20x, and $20–$100 per seat on Team. The trap. Claim emails went out around June 8, and Anthropic's support doc is explicit: the credit requires a one-time opt-in claim. "I got the email, so I'm fine" is exactly the failure mode — the email is the start of the task, not proof you finished it. Unclaimed accounts sit at $0 agent credit when the meter flips Monday, and CI jobs, cron'd scripts, and SDK agents start failing. There's a second latch. Overflow billing — what happens after the monthly credit runs out — is a separate toggle in account settings. Leave it off and exhausted accounts don't degrade gracefully; automated requests fail outright. The two actions. One: claim the credit (the link is in Anthropic's email; the support doc walks through it). Two: open account settings and answer the overflow question on purpose — pay-as-you-go past the credit, or a hard stop. Either answer is defensible. Not knowing which one you picked is not. What's unaffected. Interactive use. claude.ai chat, the desktop and mobile apps, Claude Code in your terminal, and Cowork all stay on normal subscription limits. This is about unattended, programmatic traffic only. Why it matters: this is a separation, not a price hike — Anthropic is fencing automated traffic onto its own meter. But the default state is an outage, and it lands Monday morning, exactly when weekend cron failures surface. If your team runs Claude agents, someone owns this before end of day. Ship it? → Do the two clicks today. Both of them. Claim, then toggle. Three days is plenty — unless you spend them assuming the email handled it. |
Source: Anthropic (ClaudeDevs) announcement · Anthropic support doc: use the Agent SDK with your Claude plan · The New Stack explainer 📊 Quick poll Have you claimed your Agent SDK credit yet? Claimed and toggled · Claimed, forgot the toggle · …there's a claim? Reply and tell us: did you claim yours? |
🗞️ Tour de Headlines 🧬 Following up: Hermes gets a Profile Builder — isolated work agents without the CLI ritual We covered Hermes Desktop on June 4. This week, Nous shipped the ops layer: a Profile Builder inside Hermes Agent's local dashboard (hermes dashboard, loopback-only at 127.0.0.1:9119) — one guided form for an agent's identity, model and provider, skills (including hub installs), and MCP servers. Each profile is a fully separate home directory: its own config, env, SOUL.md, memory, sessions, cron jobs, and state database. So a coding agent and a client-facing research agent share a machine without sharing state — formerly a multi-step CLI ritual, and the exact ops gap that kept open-source agents out of client work. The "agent fleet" category everyone sells as enterprise software is quietly becoming a free localhost feature. Source: Nous docs: profiles · MarkTechPost 📦 OpenClaw cuts its first monthly stable release — v2026.6.5 OpenClaw shipped stable v2026.6.5 on Tuesday — the first cut on a new monthly release train: YYYY.M.PATCH versioning, with June's floor pinned at 2026.6.5. The 30+ fixes are agent-reliability plumbing: hardened session recovery, MCP tool results coerced at the boundary so a misbehaving server can't poison session history, model reasoning stripped before channel delivery, and stricter checks on plugin installs. For the framework that opened 2026 with a one-click RCE and a malicious-skills mess, a vetted monthly version you can pin — instead of nightly roulette — is the upstream answer to enterprise doubt. And it's on npm as latest already. Source: GitHub release: v2026.6.5 · OpenClaw release policy 🌑 Gemini went dark worldwide — five days before Agentforce flips to Gemini Flash Gemini was down for at least six hours on June 10: a global outage across web, mobile, and Gemini in Chrome, on free and paid tiers, with users hitting errors 1076 and 1099. Google's Josh Woodward acknowledged it mid-incident; mitigations landed late morning Pacific. Coverage centers on the app, not the API — but the lesson holds either way. And the timing stings: Salesforce flips Agentforce onto Gemini 3.5 Flash on June 15 (our June 9 coverage). Agentforce shops just watched their future model layer take a nap. The enterprise model question is no longer capability — it's reliability. Multi-vendor failover is an architecture requirement, not paranoia. Source: TechRadar live blog · Tom's Guide live updates
🔧 Tool of the Day LiteLLM — one API in front of every model vendor Wednesday's outage answered "why." This is the "how." LiteLLM is an open-source LLM gateway: one OpenAI-format API across 100+ providers, with routing, fallbacks, retries, and spend tracking per key. The one-liner: when Gemini 500s, your agent fails over to a second vendor instead of failing. Evergreen pick — wire it in before the next fire drill, not during it. github.com/BerriAI/litellm →
🔗 Worth a Click Buy OpenAI with budget you already have. OpenAI on Oracle Cloud — OpenAI models and Codex become payable with Oracle Universal Credits "in the coming weeks." Announced, not yet available — but procurement, not technology, is the real adoption bottleneck.
⚡ Stat of the Day 276,000. That's how many employees KPMG is rolling out Microsoft 365 Copilot to, across 138 countries — with Agent 365 governing how every agent is deployed, monitored, and secured. The largest disclosed enterprise agent rollout is led by the control plane, not the agents. "Manage and audit" is what gets a 276K-seat enterprise to sign. |
The Agent Stack is a free daily brief for B2B builders. Subscribe | 
