Weekly AI Digest — March 9–15, 2026: Agents Go Enterprise, Meta De

Maya ChenAI Researcher & Product Reviewer

Welcome back to the AIStackDigest Weekly Roundup. Here’s everything that moved the needle in AI this week — from enterprise agent battles to delayed model launches, military AI controversy, and a cautionary tale about leaning too hard into generative content.

🤖 The Agentic Wave Crashes Into Enterprise

This week made it crystal clear: the AI agent wars are no longer a developer playground — they’re a full-on enterprise battleground. Three major moves underscored the shift.

Perplexity launched “Computer for Enterprise” at its Ask 2026 developer conference. The product is a multi-model AI agent capable of orchestrating 20 different AI models simultaneously, with native Slack integration and Snowflake connectors baked in. The pitch is direct: replace Microsoft Copilot and Salesforce’s Einstein AI with something faster, more flexible, and model-agnostic. It’s an aggressive play — and the timing is no accident.

Anthropic expanded Claude’s cross-app intelligence, enabling the model to share context across Microsoft Excel and PowerPoint sessions. Instead of re-explaining your spreadsheet every time you switch to a slide deck, Claude now carries the thread across both — a genuinely useful workflow improvement for knowledge workers. It’s also an interesting competitive wrinkle given that Claude partly powers Microsoft’s own Copilot Cowork, which was also announced this week.

Google upgraded Gemini for Workspace, enabling it to synthesize information across Docs, Sheets, Slides, Gmail, and Drive simultaneously to produce polished professional content in seconds. Google frames it as going from “assistant” to “collaborator.” Whether enterprise users buy that framing will depend on how well it handles messy real-world data — historically Workspace AI’s weak point.

🥑 Meta’s Avocado Needs More Time to Ripen

In a notable stumble for Meta’s AI ambitions, the company has postponed its next major language model — internally codenamed Avocado — from a March launch to at least May. The New York Times reported that performance benchmarks have fallen short of rivals, particularly Google. This is a significant delay: Avocado was supposed to be Meta’s first flagship release since the high-profile hire of Scale AI CEO Alexandr Wang, who was brought in to overhaul Meta’s AI strategy.

Meta has already spent billions trying to close the gap with OpenAI and Google. The delay raises real questions about whether the company’s infrastructure-heavy, open-weight approach can keep pace with the increasingly rapid release cadence of its rivals. In better news for Meta, the company also launched the MTIA 300 chip, custom silicon designed for training ranking and recommendation systems across Instagram and Facebook. The MTIA 400/450/500 series — targeting generative AI inference — is on the roadmap for 2026–2027.

⚔️ Military AI: Palantir’s Maven and the Ethics Debate

The week’s most sobering story came out of AIPCon, Palantir’s annual conference. A live demo of the company’s Maven Smart System — a military targeting platform — drew widespread attention after a clip went viral showing the Department of War’s Chief Digital and AI Officer demonstrating the ability to designate a strike target in three mouse clicks: “Left click, right click, left click.” The imagery of an AI-powered Kanban board for lethal targeting decisions hit a nerve across the tech industry.

This came the same week that tech workers at multiple AI labs continued voicing concern about Pentagon partnerships — a tension that’s been simmering since Anthropic’s own DOD contract disclosure earlier this month. Techdirt’s Mike Masnick published a widely-shared piece connecting the history of NSA mass surveillance to the current direction of AI militarization. Expect this debate to intensify heading into Q2.

🔬 Research Corner: Karpathy’s Autoresearch and Cooperative Agents

On the research side, Andrej Karpathy released a new open-source project called autoresearch that’s generating significant buzz. The tool enables an AI agent to read its own source code, formulate a hypothesis for improvement, modify the code, run the experiment, and evaluate the results — entirely autonomously. Karpathy suggests researchers could run hundreds of AI experiments per night, radically compressing the ML iteration cycle. Early reactions from the research community range from excited to genuinely alarmed, in the best possible way.

Separately, Google’s DeepMind team published findings showing that AI agents trained against unpredictable, diverse opponents develop cooperative strategies without explicit orchestration rules. The implication for enterprise multi-agent systems is significant: rigid workflow blueprints may be less important than adversarial training diversity. This could meaningfully influence how next-generation agentic frameworks are architected.

💼 Business and Deals

Netflix acquired Ben Affleck’s AI startup for approximately $600 million. The startup’s approach centers on AI as a creative accelerator for human filmmakers rather than a replacement, positioning it differently from other generative media plays. Netflix’s bet signals a belief that AI-assisted production will meaningfully reduce blockbuster development timelines and costs within the next few years.

Adobe CEO Shantanu Narayen announced he is stepping down after 18 years. His tenure saw Adobe navigate the collapsed Figma acquisition and a significant pivot to AI-powered creative tools under the Firefly brand. His successor will inherit a company at an inflection point: Firefly has traction, but competition from AI-native design tools is intensifying rapidly.

In startup news, Manufact raised $6.3M from Y Combinator to build open-source infrastructure for the Model Context Protocol (MCP) — increasingly dubbed the “USB-C for AI.” With both ChatGPT and Claude adopting MCP, the protocol has moved from niche developer experiment to foundational plumbing in a matter of months. Manufact is betting that infrastructure around MCP is itself a multi-hundred-million-dollar opportunity.

⚠️ Cautionary Tale: BuzzFeed’s AI Reckoning

Three years ago BuzzFeed went all-in on AI-generated articles and quiz content. This week, the company reported a $57.3 million loss for 2025, with its stock sitting at $0.70 per share. The lesson is stark: readers can tell when content is AI-generated, they don’t like it, and AI-slop doesn’t drive the engagement metrics that digital advertising requires. The remarkable part? CEO Jonah Peretti appears undeterred, announcing plans for “new AI apps.” The market has already voted.

Amazon offered a contrasting data point in AI governance. After recent AWS outages were linked to errors made by AI coding agents, the company’s eCommerce SVP called an all-hands meeting and announced that junior and mid-level engineers will now require senior sign-off on any AI-assisted code changes. It’s a pragmatic course-correction that reflects where enterprise AI governance is genuinely headed: more human oversight, not less, as autonomous coding expands.

🌍 Platform Expansions: Gemini, Copilot, and Health AI

Google expanded Gemini in Chrome to Canada, New Zealand, and India, with support for 50+ languages now live including Spanish, French, Hindi, and Chinese. The browser-integrated assistant can answer questions about on-screen content, draft Gmail messages, compare products across tabs, and remix images — making it arguably the most deeply integrated AI assistant in any mainstream browser today.

Microsoft brought Copilot to Xbox, completing its rollout across mobile, Windows 11, and handheld gaming (the Xbox Ally). Gaming is a genuine differentiator for consumer AI: an assistant that helps troubleshoot games, suggest character builds, or surface lore on demand has obvious everyday utility that productivity AI often struggles to demonstrate.

Amazon quietly expanded its Health AI agent, a HIPAA-compliant tool for answering general health questions, analyzing medical records, and connecting users with One Medical professionals. It’s the first serious competition for ChatGPT for Healthcare from a platform with actual healthcare delivery infrastructure already in place.

👀 What to Watch Next Week

Thinking Machines Lab (Mira Murati’s startup) announced a long-term gigawatt-scale partnership with Nvidia for model training. Watch for the first model or product preview — the startup has been quiet since several founding members returned to OpenAI, and this partnership signals a major compute commitment is locked in.
Meta’s Avocado: with May now the target, expect benchmark leaks and preview posts to start appearing. Any further slippage will be a significant reputational hit for Meta AI.
MCP ecosystem: with Manufact’s funding and both Anthropic and OpenAI embracing the protocol, a wave of new integrations and developer tooling is imminent. This could become the week’s biggest story two weeks from now.
AI governance and military AI: the Palantir Maven demo has put AI ethics squarely back on the front page. Congressional interest is growing, and internal pressure at AI labs over government contracts is at a high point. Watch for official statements or policy announcements.
Adobe’s leadership announcement: Narayen’s successor will signal which direction Adobe intends to go — full AI-native transformation or cautious integration. The industry is watching.

That’s the week in AI. The enterprise agent wars are real, the model race continues with notable stumbles, and the ethical reckoning around military AI is no longer a background conversation. See you next Sunday. — AIStackDigest

What to Read Next

Bookmark aistackdigest.com for daily AI tools, reviews, and workflow guides.

This article was produced with the assistance of AI tools and reviewed by the AIStackDigest editorial team.

Weekly AI Digest — March 9–15, 2026: Agents Go Enterprise, Meta Delays Avocado, and the AI Arms Race Heats Up