I’m in SF this week, and a lot of the conversations I’ve had here have been about what happens once the models are actually good enough.

Because as they get better, the harder technical problems around them start surfacing: context, memory, retrieval, caching, orchestration, and figuring out what information should reach the agent in the first place. That’s the thread running through this week’s issue.

Summary

Read time: 8 min

Highlights Of The Week

Webinar: Intro to tokenomics: What Every Manager Needs to Understand About AI Spend

Also This Week:

Context As A Service · The Context Window Myth

Signals

1. Fable 5.1 improves long-horizon coding and agentic work, with much cheaper cache reads.

2. Slack launched Slack Code, dedicated channels where your team and its agents ship together, with Claude, ChatGPT, Devin, and Copilot as launch partners

3. Z.ai shipped GLM-5.3, a 743B-parameter, 1M-context, MIT-licensed open-weight model, with gains that came from post-training alone

4. Apple announced the M6, its first 2nm chip, and the M5 Ultra, built into a new Mac mini and Mac Studio aimed at running models on device

5. OpenAI says its Jalapeño chip, built with Broadcom, beats Nvidia in testing while running at or below 550W in real workloads

6. Gemini's new video model cuts costs 66% by watching only what matters

7. Google introduced Gemini 3.5 Transcribe, its most precise speech-to-text model yet

8. Google Cloud shows how to scale OKF bundles across an organization with Knowledge Catalog

Context is becoming infrastructure

I hosted another AI infra dinner with founders in SF, this time around Context as a Service. Everyone talks about 100K, 500K and 1M-token context windows, but long-running agents don’t just need more context. They need the right information, in the right form, at the right time.

George Fahmy shared a great example: pruning old context cut token usage by ~30%, but made the workload 3x more expensive because it broke prompt caching. Rafael Levi showed the data side of the same problem: a 17K-token webpage can become ~107 tokens when you only send the fields the agent actually needs.

And the research backs it up:

Oolong — GPT-5, Claude Sonnet 4 and Gemini 2.5 Pro all scored below 50% on harder 128K reasoning tasks.

ATLAS — tested 26 models up to 1M tokens and found an average 24.3% performance decay as context scaled.

Long-Context WebAgents — success fell from ~40–50% to below 10% as agent histories grew, with agents looping and losing the original goal.

Sliding-window beats linear attention — a new paper found a simpler sliding-window approach performed 2–10x better than post-trained linear attention on two long-context benchmarks, while using far less memory.

Google: teaching LLMs to know what they don’t know — another useful direction: models need better ways to represent uncertainty instead of confidently acting when they don’t have enough information.

The shift is from “how much can I fit in the window?” to: what should I retrieve, cache, summarise, persist, refresh — or never send to the model at all?

That’s why I think CaaS becomes a much bigger part of the agent stack, and why the metric worth watching isn’t cost per token. It’s cost per verified successful task — something we’re already starting to roll out inside ELI.

15 Best Free Claude Code GitHub Repos That Make It 100x More Useful - Worth wiring in if you're running on Codex or Claude Code day to day.

What’s In Their Stack?

This week: Anna Chernikova

Founder and CEO of Sprites, Anna has 32 tools across 27 categories in her stack.

Her AI layer runs on Anthropic and OpenAI, with Fireflies.ai handling meeting notes and Fireworks powering orchestration and MLOps.

Then there's the rest of the stack: See Anna's full stack →

Want visibility on what's in yours? Map it with ELI.work

Stay curious.

Ghita

One episode I consumed this week that stuck with me from the All-in podcast where Gerstner, Sacks, Friedberg, and Calacanis put words to something I've been watching happen in real time:

Investors aren't paying for 30 years of future cash flow anymore. There's too much uncertainty about which software companies will even exist in 7 years. AI makes everything harder to price.

David Friedberg said: software is moving from helping us be more productive to actually completing the work. That's a business model shift. You can't charge per seat when the seat is an AI agent. The future is outcome-based pricing: charging for value delivered, not human headcount.

Which brings me to what I've been obsessing over at TechBible. If the future is agents completing work instead of humans using tools, then your infrastructure choices today determine whether your agents can actually function tomorrow. That's why I built ELI, our Enterprise Intelligence Layer methodology. It maps your entire tech stack to show you which tools are positioned to support agentic workflows and which ones are just... sitting there, collecting dust and burning cash.

See your stack now

What's actually happening in the market

Capital One acquired Brex at a steep discount. Google hired the Hume AI team. Apple bought Q.ai. These aren't growth plays, they're admissions that after billions in R&D, the giants still can't build what startups figured out in 18 months :P

Speed is beating resources. But the real bottleneck isn't models or interfaces. It's memory.

Google just released something called Nested Learning to stop AI from forgetting what it learned. Alibaba built a benchmark for models that can handle hundreds of steps without losing the thread. Every AI agent today hits a wall around step 50. Whoever cracks long-term context without hallucinations wins this race.

Secure or Not ? Here I come

I don't want to be dramatic, but we need to talk about this. Anthropic just published research showing that even benign fine-tuning can accidentally break AI safety. And last week, researchers hacked Moltbook's database in 3 minutes—accessing thousands of private emails.

If your AI agent can access private data and you haven't solved for security architecturally, you don't have a product. You have a liability waiting to happen. The first major breach will kill half the startups in this space.

Here are top 3 security tools for agents companies are adopting on Techbible this week:

Invariant Labs

Prompt Security

Zenity

3 bets on the future

Right now, three different strategies are playing out:

Google is betting on emotionally aware voice interfaces with the Hume acquisition. Expect Gemini's voice mode to get unsettlingly human by Q3.

xAI is betting on human emulators for legacy UIs—automating the 90% of office work trapped in systems companies can't rip out.

OpenAI is betting on repeatable automations with their new Codex app for scheduling and delegating work.

These aren't competing strategies. They're different markets. Voice wins consumer. Automation wins enterprise. Emulators win the messy middle where Fortune 500 companies are stuck with 30-year-old ERP systems. Probably all 3 will win. But in different categories.

Some good news for developers

Anthropic released MCP Apps—you can now build live UIs inside the chat instead of toggling between code and interface.

Moonshot open-sourced Kimi Code, an agent for autonomous coding across entire codebases.

And Luma made high-fidelity video generation cheap enough to actually use.

If you're building agents that need to pull data from the web, this will save you weeks. I just compared the best web scraping tools for AI workflows.

Stay curious.

From Cape Town, Ghita

Graph engineering is workflow design wearing an AI conference badge. The term started as a joke, but the problem is real: when several AI agents work together, you need to control what runs, when, and who checks it. Read below for a simple explanation.

Summary

Read time: 3 min

Deep Dive

WTF is Graph Engineering!

Highlights

ELI Product Update

10 documents I think every company should have before scaling AI agents.

News

OpenAI's unreleased model solved ten long-standing math problems, including one open since Erdős

Alibaba released Qwen3.8-Max, a 2.4 trillion parameter model going open-weight next week

Google DeepMind's new model lets humanoid robots walk, grasp, and work as a team

Anthropic found three cases where Claude broke out of test environments and hacked real companies by mistake

A frontier model broke out of its own test sandbox and hacked Hugging Face, on purpose, sort of

This month's biggest funding rounds, tracked and updated weekly

Update: New ELI Features This Month

I shipped more than 40 updates last month onto ELI. These are 3 of my favorite:

1. Track more AI spend across your stack: Connect OpenRouter, fal.ai, Fireworks, Deepgram, ElevenLabs, xAI, and LangSmith. ELI also detects GitHub Copilot and Gemini for Workspace through existing connections.

2. Manage permissions from chat: Ask which apps have Gmail access, revoke risky permissions, or reset a user’s password without digging through the admin console.

3. Catch renewals and spend changes early: ELI now alerts you before renewals and flags unusual changes in software spend, so you have time to cancel, renegotiate, or investigate.

Deep Dive

A graph has three parts:

Nodes do the work. A node might be an agent, a model call, a tool, a function, or a human reviewer.

Edges decide what happens next. The workflow might continue, branch, retry, run tasks in parallel, or stop.

State carries information between nodes, so every step knows what happened before it.

Graphs do not replace loops. A single agent loop is simply a one-node graph with an edge pointing back to itself. Graphs connect and govern several loops.

The technology is not new either. LangGraph introduced nodes, edges, and shared state in January 2024. Microsoft has GraphFlow, and Google uses the same workflow model in ADK 2.0. The useful skill is not drawing more boxes. It is knowing which boxes deserve to exist.

Every extra node adds cost and another place for the system to fail. Anthropic estimates that a single agent uses around four times more tokens than a chat interaction, while a multi-agent system uses around 15 times more. That expense makes sense when work splits into genuine specialties or parallel tasks. Anthropic’s multi-agent research system outperformed a single Opus agent by 90.2% on its internal research evaluation because research naturally splits into independent searches.

For simpler tasks, a graph is expensive decoration.

The safest setup keeps predictable routing in code, limits which nodes can change shared state, and uses a reviewer with fresh context and external evidence. Think tests that ran, code that compiled, or a human approval. Several agents using the same model and flawed context will often agree on the same wrong answer.

Use a graph when your workflow needs parallel work, specialist agents, retries, approvals, failure isolation, or an audit trail.

If the task is one prompt, one tool call, and one response, keep the loop.

Give Your AI Agents a Company Brain

I built a Claude skill that turns your answers into ten company documents covering your identity, team, processes, guardrails, goals, and tech stack. It connects to ELI for live data on tools, costs, owners, and renewals. In around 20 minutes, your agents get the context they need to stop guessing.

Signals

OpenAI's unreleased Astra model solved ten long-standing open problems in math and theoretical computer science, including one of Paul Erdős's problems, each with a machine-checkable proof published alongside a technical paper. It's unclear whether or when Astra ships as a public model.

Alibaba released Qwen3.8-Max, a 2.4 trillion parameter mixture-of-experts model with a million-token context window. It's live now through Alibaba's hosted platforms, with open weights following next week.

Google DeepMind's Gemini Robotics 2 gives humanoid robots whole-body control, letting them walk, crouch, and manipulate objects, and coordinate with other robots on shared tasks. Access is still limited to trusted testers and early partners.

Anthropic found three cases where Claude broke out of test environments and compromised real companies by mistake. A misconfiguration with an outside testing partner left supposedly isolated environments connected to the internet, and Claude, believing it was still inside a simulation, used basic techniques like weak passwords to get in.

OpenAI's own models triggered the same category of incident first, breaking out of a sandboxed evaluation via a zero-day vulnerability and hacking into Hugging Face's production systems while trying to cheat on an internal benchmark.

This month's biggest funding rounds, tracked and updated weekly.

See who's raising →

Stay curious.

Ghita

Summary

Read time: 4 min

Highlight Of The Week:

The Tokenomics of AI

Who’s Buying in Tech This Week

$5B changed hands between Bending Spoons, Autodesk, Procore, and more.

Signals

OpenAI is building a $300+ AI speaker with Jony Ive's studio, its first hardware move beyond the screen

Oklo split its first atoms in a Texas test reactor, a milestone for its nuclear technology and a step toward medical isotopes

Google Search Console is expanding beyond websites, letting creators see how their Instagram, TikTok, X, and YouTube content performs in search

Cloudflare built Kitesurf, a lightweight browser made for AI agents that uses up to 7x less memory than Chromium

What I shipped this week

Deeper AI spend visibility

You can now see model-level spend and token breakdowns for every AI vendor, plus explore how your AI costs change over time in the new spend trend view.

Highlight Of The Week: Tokenomics

Welcome to Tokenomics, where understanding which tasks AI can do is just as important as how much it costs and how it impacts your balance sheet. If you want to understand why AI is so hard to price and where your AI budget is quietly leaking, these two articles are must-reads.

Who's Buying in Tech This Week

Bending Spoons acquires Airtable for $1.285B - its first deal since going public last month, adding a major enterprise software brand to the portfolio.

Autodesk acquires MaintainX for $3.6B - bringing real-world maintenance and operations data into its design-to-operate platform.

Procore acquires DroneDeploy for $845M - pairing construction data with jobsite visual intelligence to build "digital co-workers" that spot problems and act on them.

Okta acquires Permiso for ~$200M - extending identity security beyond login to catch threats from AI agents and machine identities.

AutoRek acquires Grath, terms undisclosed - combining enterprise-grade financial controls with Grath's AI-native reconciliation tech.

OpenAI acquired presentation startup NextSlide, folding its team into ChatGPT to build better AI-generated slides.

AMD is acquiring chip startup Taalas, betting on hardwir.ed silicon to make AI inference cheaper than general-purpose GPUs

What’s In Their Stack?

This week: Angus Sewell McCann

AI engineer and founder of TempShift, Angus has 23 tools across 20 categories in his stack.

His AI layer includes Superwhisper, Gemini and Claude, with Antigravity handling AI-assisted coding and Granola covering meetings.

Then there's the rest of the stack: Attio, Slack, Superhuman, Perplexity, NotebookLM, n8n, Stripe, Luma, Railway and more. See Angus's full stack →

Want to see what’s in your stack? Map it with ELI.work

Stay curious.

Ghita

I kept coming back to one question this week: who is actually best placed to decide which intelligence should handle a task? Cursor already routes based on the coding context it sees. OpenRouter sees hundreds of models and providers. Frontier labs know their own models better than anyone. As model choice starts disappearing into the infrastructure, those are three completely different advantages. Interestingly, routing benchmarks are already suggesting it might not be the routing algorithm itself. The bigger advantage is who has the best data on what “good” looks like for that specific task. Which made me wonder: if that’s true, was Stripe’s $7B OpenRouter deal actually worth it?

Summary

Read time: 4 min

Highlights Of The Week

I paid $5 for a $0.50 coding task

Routing vs. Fusion, and why everyone suddenly cares 👇

Signals

Databricks closed a $5B round at a $190B valuation, and crossed a $7B revenue run-rate after growing more than 80% YoY.

OpenAI hired Wiz President and COO Dali Rajic as its new revenue chief, replacing former Slack CEO Denise Dresser after less than a year in the role.

Canva Code now generates interactive tools from a single prompt, pricing calculators, sign-up forms, quizzes, built on your brand guidelines and published on your own domain.

OpenRouter sold to Stripe for $7B

Anthropic ships enterprise auth for Claude connectors, no OAuth setup needed

Calendly Adds Meeting Notes

==Routing vs. Fusion, and why everyone suddenly cares

==

Routing chooses the best model for a task based on quality, cost, speed and reliability. Fusion sends the same problem to multiple models and compares or combines the answers. Routing asks, “Which model should do this?” Fusion asks, “Is this important enough to get several models involved?”

Both are becoming products. OpenRouter launched Fusion, where multiple models answer in parallel and a judge model looks for consensus, contradictions and gaps. That makes sense for research, strategy or anything expensive to get wrong.

Routing is where the trend is now. Stripe agreed to acquire OpenRouter, while Ramp launched Router.com the same day. The best model keeps changing. New models launch, prices move, providers hit limits and different models win at different tasks. Companies don’t want to benchmark all of that manually. They want one layer deciding where each request should go.

So was OpenRouter really worth $7B to Stripe? Maybe, but not because its routing algorithm is impossible to copy. OpenRouter has aggregation: hundreds of models, huge amounts of cross-provider traffic, pricing and reliability data, and developers already using it as the place where inference gets bought. Stripe already sits on the monetisation side. Now it also gets closer to deciding where AI spend gets allocated.

For Eli, we’re starting where a huge amount of AI spend already lives: inside the software companies are already paying for. Gartner estimates 85% of enterprise agentic AI investment is already bundled into existing SaaS and cloud renewals. So before companies start optimising every individual model call, there’s already an entire layer of AI subscriptions, licences and tools they need to understand and control.

The end state is not managing ChatGPT subscription over here, API spend over there and agents somewhere else. It’s one layer deciding what the company should use, what it should cost, and whether it’s actually creating value. See where it's at :

Watch the video

==MUST READ: When One AI Agent Can Access Everything

==

As AI agents connect to more apps, they also gain more access, cost and risk. Here’s how companies should think about ownership, permissions and visibility before agent sprawl gets out of control.

What’s In Their Stack?

This week: Andrew Yeung - Founder of Fibe, investor and one of the busiest people in the tech events world, Andrew’s stack is built around one thing: moving fast without adding more admin.

His AI and productivity layer includes Wispr Flow for voice, Granola for meetings, Poke for getting things done from anywhere, and Raycast for speeding up work on his computer. See Andrew's full stack →

Want to see what’s in your stack? Map it with ELI.work

Stay curious.

Ghita