Suprvisr TLDR - May 11 2026

We are super excited to be announcing that Suprvisr AI is shipping long-duration agents next week. Hours-long research projects, document review at volume, monitoring systems for the changes that actually matter. They keep working while your team is doing other things, with a full audit log of every step.

Agents are meant for tasks that require planning, careful data orchestration, reasoning and version iterations as well as human in the loop cycles with agentic confidence scoring of retreived data.

We're launching a single researcher agent that can curate powerful competitive analyses, market research, go-to-market strategies or other complex data-backed research - stay tuned as we create more capable agents that can actually contribute to your organization.

(no credit card required)

At its developer conference on May 6, Anthropic announced a deal to use all of the capacity of SpaceX's Colossus 1 data center in Memphis. That's more than 300 megawatts of compute and over 220,000 Nvidia GPUs, online within a month. Claude Code's rate limits doubled the same day.

Three months ago Elon Musk publicly called Anthropic "woke," "misanthropic," and "evil." This week he posted that the team didn't "set off [his] evil detector." Anthropic, for its part, is paying its loudest critic for the only thing it can't get fast enough anywhere else: power. Both companies also flagged interest in orbital data centers, which sounds absurd until you remember launches are now cheaper than grid upgrades in some places.

Anthropic has signed compute deals with Amazon (5 gigawatts), Google and Broadcom (another 5), Microsoft ($30 billion of Azure capacity), and Fluidstack ($50 billion in infrastructure spend) in the past two months. Access to SpaceX's Colossus 1 data center is strategic to maintaining their current momentum.

Now I can go back to building "Pawtify" - An AI-curated Spotify playlists for dogs, that matches music to their tail-wag BPM unhindered by my usage limits.

Anthropic's annualized revenue, up from $9 billion at the end of 2025. CEO Dario Amodei told developers this week that Q1 growth came in at 80-fold annualized, roughly eight times what the company had planned for. He called it "insane and unmanageable" on stage. Demand outran infrastructure, Claude users hit reliability and rate-limit issues for weeks, and the SpaceX deal you read about above stopped being a strategy and started being triage. Every week of constrained capacity is a week competitors close the gap.

At the same conference where it announced the SpaceX deal, Anthropic introduced a research-preview feature called "Dreaming," a scheduled process that reviews an agent's past sessions, extracts patterns, and curates its memory so it gets better at the job over time. Two related features shipped the same week for Claude Managed Agents: outcomes-based grading (a separate grader agent scores work against your success criteria, with no shared context to prevent gaming) and multi-agent orchestration. Memory updates can be auto-applied or held for human review.

The bigger story for us is that agentic AI is moving from a tool you use transactionally to a tool that gets better the more you integrate it into your core workflows. Every hour your team spends with it makes the next hour more valuable, which is a different kind of return than most AI products deliver. It also raises a governance question worth asking before you delegate anything important: who reviews the memory updates, and how do you see what changed. We're seeing some companies adopt git-style agent management like Cloudflare that seems interesting for tracking agent harness changes and observability.

ChatGPT got smarter and stopped relying on cute emojis. On May 5, OpenAI replaced GPT-5.3 Instant with GPT-5.5 Instant as the default model behind ChatGPT, claiming a 52.5% reduction in hallucinated claims on high-stakes prompts in law, finance, and medicine. Responses are tighter, the conversational tone is dialed down, and a new "memory sources" feature lets users see which past chats shaped a personalized answer. If your business operates in a regulated industry, that 52.5% number is the one to bring to your next compliance review. Worth retesting your team's top three recurring prompts either way.

OpenAI's coding tool developed a goblin obsession. OpenAI published a post-mortem this week explaining why GPT-5.5 in its Codex coding tool kept inserting fantasy creatures into responses. A retired ChatGPT personality called "Nerdy" had rewarded playful language in training, and the model learned that goblin-mentions got rewarded. Goblin references jumped 175% after GPT-5.1 launched and got worse from there. Codex now ships with a system prompt that tries to ban creature words. It should reinforce to your team, the LLM you use contains millions of small reinforcement decisions, most of which nobody at the lab made consciously. Spot-check what's going to your customers, or you could be emailing them about goblins.

TELUS got caught using AI to mask offshore agents' accents. The Globe and Mail reported that TELUS Digital is running overseas call-centre agents through real-time AI accent-softening from a vendor called Tomato.ai (now sanas.ai). TELUS markets the technology on its own site as "AI-powered speech enhancement" that "softens accents" while "preserving voice identity." One union representative told MPs that an offshore agent had been seen flipping the masker on and off mid-call to demonstrate it. Bell and Rogers explicitly told the Globe they don't and won't use the technology. Unifor called the practice deceptive at a parliamentary committee, asking for mandatory disclosure when AI is in the call. The disclosure question is now in front of Canadian regulators in plain sight. Thanks Telus!

This week, we'd tell you to spend ten minutes on the disclosure question before someone else does it for you. Pick one place in your business where AI talks to a customer, an employee, or a candidate. Three questions: When does the person on the other side learn an AI is involved? If they ask directly, does your tool answer accurately? Would a typical customer be surprised, in a bad way, to find out what you're doing? If any answer is uncomfortable, fix it before a journalist or regulator finds it for you. The work is figuring out what you're comfortable saying out loud, then building the workflow to match.