AcceleratingBiz®

Growth CEO’s AI Briefing · Edition #35

Agents Take Action

Thursday, May 28, 2026 · 3 Developments · 2 Perspectives · 2 Case Studies · 1 Resource

Designed by AcceleratingBiz®. Curated, synthesized, and produced by Claude agents. AI-assisted production means occasional errors may appear.

A single Wednesday news cycle reframed what agents are for. Yesterday they suggested; today they trade stocks at Robinhood, file tax returns at 97% accuracy, raise billions at decacorn valuations, and ship into BMW and Airbus production lines. The companion signals are about who controls them, who pays for them, and who breaks them.

The strategic shift is no longer about deciding whether to deploy agents. It is about the permissions, audit trails, and recovery paths you put around them before someone else's agent transacts with yours.

Developments

Business

Cognition closes $1B at $26B with $492M ARR, becoming the largest independent agent lab — May 27, 2026

Cognition announced May 27 that it raised more than $1 billion at a $26 billion post-money valuation, co-led by Lux Capital, General Catalyst, and 8VC, with Ribbit Capital, Atreides, and Founders Fund participating. The company disclosed $492 million in annualized run-rate revenue and enterprise growth of more than 10x since January, citing customers including Mercedes-Benz, Itaú, Exa, and Modal. Devin reportedly writes 89% of Cognition's own code. AINews framed the moment plainly: "officially the largest remaining independent agent lab in AI," and one of the clearest signals that coding agents are converting into large enterprise businesses. The valuation is 2.5x its $10B September Series C.

So what

The agent-coding category now has its consolidating leader, which collapses the "build vs. buy" calculus for any growth company running its own developer-productivity initiative. If your CIO is still piloting two or three agentic-coding vendors, this is the moment to consolidate to one and negotiate enterprise pricing while the category is funded but not yet locked-in. The bigger strategic question for non-software CEOs is what your industry's equivalent of Devin looks like, and which Cognition-tier company will absorb the margin in your vertical first.

⚠ Hype check: The $492M ARR and 89%-self-written-code figures are company-disclosed pre-audit. AINews is now owned by Cognition, which the newsletter discloses; weigh AINews coverage of Cognition accordingly.

Confidence: Established fact — round confirmed by TechCrunch and Cognition's own announcement; ARR and growth figures remain company-disclosed.

Go deeper Primary source (TechCrunch) · Curated (Cognition Series D blog)

Customer Behavior

Robinhood gives AI agents a brokerage account and a virtual credit card, both wired via MCP — May 27, 2026

Robinhood launched on Wednesday a beta program letting users create a dedicated AI-agent account, fund it from their main wallet, and connect external agents through open MCP servers to read portfolios, run analysis, and execute trades. The agentic Gold Card extends the same pattern to purchases, with monthly limits and approval workflows. Stocks come first; options, crypto, futures, event contracts, and prediction markets are scheduled to follow. The Neuron framed the editorial implication directly: "Agents are moving from 'help me think' to 'act on my behalf,'" which means winning agent apps now need "permissions, spending caps, audit logs, rollback buttons, and a panic switch."

So what

Robinhood just published the consumer-facing template for transactional-agent UX, and that template will become the default expectation for B2B agent design within two quarters. If your product accepts agent traffic, treat MCP as the inbound API of record and start designing the audit, rollback, and spending-limit UI now. American Banker calling it a "wake-up call for banks" is the right read — the moat in agent commerce migrates from rails to permissions architecture.

Confidence: Established fact — Robinhood newsroom announcement plus TechCrunch and WSJ primary reporting; beta scope and rollout cadence as published.

Go deeper Primary source (TechCrunch) · Curated (WSJ)

Technology

Trajectory raises $15M to make continual learning a category, with Clay, Harvey, Decagon, Mercor, and Rogo as design partners — May 27, 2026

Trajectory exited stealth Wednesday with a $15 million seed from Conviction, Bessemer, Radical, and angels Jeff Dean and Fei-Fei Li. Its product turns post-launch product usage into reinforcement signal that post-trains the customer's agentic model on a weekly cadence. The launch design partners are some of the highest-signal AI-native companies in market: Clay, Harvey, Decagon, Mercor, and Rogo. The founding team draws from DeepMind, OpenAI, Apple, Meta Superintelligence, Amazon AGI, and Scale AI, with product talent from Stripe and Figma. The thesis is that the model behind a production agent should be a moving target trained on your data, not a frozen base your prompt engineering tries to coax into the right answer.

So what

"Continual learning" is the operational answer to the brittle-agent problem most CIOs are now hitting six months into deployment. For any AI-native or AI-extending growth company, this category is the difference between an agent that decays in production and one that compounds. Add continual-learning capability to your roadmap criteria when choosing the next agent vendor, and ask explicitly how the model improves from your data without leaking it back into a shared base.

Confidence: Established fact — named investors and named design partners on the record via Wired and the company.

Go deeper Primary source (Wired) · Curated (Trajectory)

Perspectives

AI Strategy & Transformation

OpenAI Foundation commits $250M to AI's economic fallout the same week Altman walks back his "jobs apocalypse" warning — May 27, 2026

The OpenAI Foundation announced Wednesday a $250 million Economic Futures initiative covering three priorities: independent measurement of AI's labor and economic effects, retraining and transition support for affected workers, and longer-term mechanisms for sharing AI-generated wealth such as tax shifts from labor to capital and "durable claims" for citizens. Days earlier, Sam Altman publicly softened his prior warning of a near-term "jobs apocalypse," telling Time he was "pretty wrong" about the speed of labor-market disruption and that the "human part" of work cannot be replaced. The timing is conspicuous given OpenAI's reported preparation for a Q4 2026 IPO at a roughly $1 trillion valuation.

So what

The largest commercial AI company is now formally underwriting the AI-displacement narrative it spent the prior two years either provoking or denying, which raises the bar on what growth-company CEOs say publicly about AI-attributed workforce decisions. Audit your AI-attributed layoff language for the next twelve months: anything that pairs "AI is doing the work" with "we're eliminating roles" is now sitting in a regulatory and reputational landscape where OpenAI itself is funding the counter-narrative. The smart posture is workforce transition framing first, productivity-gain framing second.

⚠ Hype check: Apply heightened skepticism to executive statements that reposition AI's labor impact when a $1 trillion IPO is in the regulatory pipeline. The Foundation commitment is real; the framing serves more than one purpose.

Confidence: Established fact — Foundation announcement, NPR and Time primary reporting; the interpretation that this is IPO-positioning is informed opinion.

Go deeper Primary source (OpenAI Foundation) · Curated (NPR)

AI Strategy & Transformation

Cisco: every frontier model fails under multi-turn attack, with success rates as high as 88.3% — May 27, 2026

Cisco's AI threat-intelligence team published a new evaluation Wednesday across 15 closed frontier models from OpenAI, Anthropic, Google, Amazon, and xAI, covering roughly 30,000 single-turn prompts and nearly 7,000 multi-turn attacks across 1,400 conversations. Single-turn attack success rates ranged from 2.19% to 64.91%; multi-turn rates climbed to 7.89% to 88.30%. xAI's Grok 4.1 Fast non-reasoning hit 88.3% multi-turn (up from 34.1% single-turn). Google's Gemini 3 Pro: 73.3% versus 18.1%. Anthropic's Claude Opus 4.6, the best on single-turn at 3.6%, still degraded to 16.2% multi-turn. The five families of attack that drove the failure modes are role-play, contextual ambiguity, refusal reframing, information decomposition, and crescendo escalation.

So what

Every multi-step agent your team is deploying is exposed in exactly the dimension this research now quantifies. The standard procurement question, "which model has the best safety benchmark," is the wrong question; ask instead what runtime guardrails, conversation-length limits, and red-team protocols the vendor or your own platform team will run against multi-turn prompts. Tee up a Q3 internal red-team exercise against your two most-used agents using the five attack families Cisco names; the results will reset your guardrail roadmap.

Confidence: Established fact — Cisco's full methodology and per-model results are published; replication by independent labs is forthcoming.

Go deeper Primary source (Cisco AI blog) · Curated (The Deep View)

Case Studies

Operations & Finance

OpenAI, Thrive, and Crete file 7,000 tax returns with a Codex agent that improved itself from 25% to 97% accuracy in six weeks — May 27, 2026

OpenAI, Thrive Holdings, and accounting network Crete Professional Alliance described a co-built tax-preparation agent that uses Codex to turn every practitioner correction into an automated improvement signal. The pilot processed more than 7,000 1040 and 1041 returns across 30+ accounting firms. Accuracy moved from 25% at launch to 86% at six weeks and to a reported 97% over the full pilot. Preparation time dropped roughly one third; throughput rose roughly 50%. The architectural pattern that matters: the agent reads accountant edits, generates evals from them, and ships its own pull requests back into the model. OpenAI is positioning this as the reference design for self-improving vertical agents inside professional services.

So what

The transferable pattern is corrections-as-evals, not tax specifically. Any growth company with high-volume practitioner workflows (claims adjudication, contract review, compliance checks, support triage) can build the same loop: capture the human override, route it into an eval suite, and let the agent re-train against it. Have an engineering lead sketch where in your highest-margin workflow this loop fits before the next OKR cycle; the firms that institutionalize this pattern by year-end will be the ones quoting Crete's accuracy curve in their fundraising decks.

Confidence: Established fact — figures published by OpenAI with named partner firms; outside-firm validation of the accuracy curve is not yet available.

Go deeper Primary source (OpenAI) · Curated (The Neuron)

AI Strategy & Transformation

Mistral launches "Mistral for Industrial Engineering" with Airbus, BMW, EDF, and CMA CGM as deployment partners — May 28, 2026

At its first annual conference in Paris on Thursday, Mistral AI launched a vertical agent stack pitched at heavy industry: design, simulation, quality control, and physical-AI workflows, with frontier models customized on customer drawings and blueprints and hosted on private bare-metal infrastructure. Airbus signed a five-year deal across aircraft, helicopter, and space programs. BMW Group is the central partner for its "Large Industry Model" on production lines. EDF and shipping group CMA CGM round out the launch cohort. The product is enabled by Mistral's recent acquisition of Vienna's Emmi AI and follows its earlier Mistral for Finance. New 10MW inferencing capacity opens this quarter at Les Ulis with a 1GW target by 2030.

So what

Industrial AI now has a credible European stack pitched at sovereignty-sensitive customers, which means RFPs from heavy-industry buyers (defense, automotive, energy, logistics) will begin demanding non-US foundation-model alternatives in the second half of 2026. For growth companies selling into European industrial procurement, add Mistral compatibility to your AI integration roadmap. For US-based growth companies, the deeper signal is that vertical AI stacks (industry-first, model-second) are starting to win flagship logos away from the horizontal foundation labs.

Confidence: Established fact — Bloomberg primary reporting plus on-the-record customer announcements at Mistral's Paris event.

Go deeper Primary source (Bloomberg) · Curated (The Deep View)

Resources

Tools

OpenAI ships Secure MCP Tunnel: connect private MCP servers to ChatGPT and Codex over outbound-only HTTPS — May 28, 2026

OpenAI released Secure MCP Tunnel this week, giving enterprises a way to connect internal MCP servers to ChatGPT Enterprise, ChatGPT Business, and Codex without opening any inbound firewall ports or exposing servers to the public internet. A tunnel client establishes outbound HTTPS to OpenAI; OpenAI then routes MCP traffic back through the same connection. The package complements broader enterprise plumbing in this week's Codex updates, including Workload Identity Federation and an expanded Admin API. Cumulatively, it removes the most common security objection growth-company CIOs have raised against bringing internal data sources into agent workflows.

So what

The "we can't connect ChatGPT to our internal systems because we'd have to expose them" objection is closed. If a CIO or CISO has stalled an OpenAI rollout on inbound-firewall grounds, escalate the Secure MCP Tunnel docs to them this week. The follow-on architectural question is which internal MCP servers you stand up first; the right answer is the workflow with the highest knowledge-worker hours per week, not the loudest internal proponent.

Confidence: Established fact — OpenAI developer documentation published; corroborated by TLDR AI and AINews coverage.

Go deeper Primary source (OpenAI docs) · Curated (AINews)

Sources in this edition: AINews (swyx), TLDR AI, The Rundown AI, The Neuron / Mindstream, The Deep View, Superhuman AI

← Latest Editions and Search