The task goes to the trusted
The task goes to the trusted
The Angle Issue #326
The task goes to the trusted
Lately I've had versions of the same conversation with folks in Europe and the US. The discussion starts with agents and almost always ends up being about trust.
At a recent dinner with a group of tech execs, I asked the room a question: What's a killer example of a personal agent actually making your life better? I expected a flood of stories. Instead, there were only a few. One came from an exec whose Instinct agent had sent a scheduling email on their behalf. Useful, for sure. But ChatGPT or Claude could have done exactly the same thing. They hadn't because this person didn't trust them enough to give them access to Gmail.
The other examples followed the same pattern. The agent wasn't doing much that a general-purpose assistant couldn't do. The person had decided to trust it, and the trust, not the technology, made the outcome possible.
This is the most important dynamic in AI right now: the task goes to the system the user trusts. I'm seeing it in three places.
Software Development
A few weeks ago, I wrote that code is a lot like math: either it compiles and the tests pass, or it doesn't. I said that thoughtfully constructed tests were a topic for another day. That day has arrived. More teams are running fully automated software development where nobody reads the code. They define precisely what they want to build and what the output should be. Because they control both, they don't need to look at what's in between. Agents write the code, test it, and fix the bugs. When something goes wrong, the team treats it as a problem with the agents, not a problem for a human engineer to solve.
StrongDM is the best-known example. Its AI team works under two rules: "Code must not be written by humans" and "Code must not be reviewed by humans" (Simon Willison). Addy Osmani explains the logic in The Code Nobody Reads: Trust in unread code depends on how cheaply you can check the work and undo a failure after it ships. It doesn't always work; HumanLayer ran a fully automated setup in 2025 until an outage sent the team on weeks of archaeology through code no human had ever read. Where the tests are trusted, though, the agents get the job.
Personal Agents
Products like Muse and Instinct deserve some of the excitement around them. But, as the dinner showed, their advantage so far is largely the access users have granted them rather than capability. The more interesting question is how they earn that access. They ask for it in exchange for value the user sees right away. For example, an agent asks for Gmail access so it can find recurring subscriptions the user rarely uses and cancel them. The user saves money in the first session and the agent now has access for everything that follows.
For founders, that's good news and bad news. Trust is a real advantage, so earning it early is the first job. When trust is lost, however, there may never be another chance to recover it.
B2B
This is where I spend most of my time with founders. Many of the companies I talk to are deciding whether general-purpose agents should be allowed to access their products directly, and if so, how much. Some lean toward open access on the basis that it doesn't matter whose interface the customer uses as long as the problem gets solved.
It matters considerably. If users go to another agent to complete a task, you've lost their trust to be responsible for that task. You become a backend that other agents use (or don't use!) and you lose the context you need to do this and other jobs well.
We've already seen early versions of this fight. In 2025, Salesforce changed Slack's API terms so that apps like Glean could no longer store Slack data long term. Amazon sued Perplexity over its Comet browser shopping on Amazon on users' behalf and won an injunction in March 2026. In August, the Ninth Circuit overturned it, holding that it's the user, not Perplexity, who is accessing Amazon. Locking agents out may not be an option. You have to win.
How B2B Companies Compete
Be accountable for the outcome. A general-purpose agent can't be responsible for your customer's result, but a specialized agent can. Done well, it holds the context the general-purpose agent never sees: the workflow state, the history, the edge cases that were never written down. That context lets you act proactively, which saves the customer time and leaves them less reason to ask another agent. It also gives you the confidence to offer guarantees, move to outcome-based pricing, and own mistakes when they occur. This is how you earn trust rather than ask for it. The personal agent playbook applies here too: Deliver visible value early and the access follows.
Be the specialist the generalist calls. Some users will always start in ChatGPT or Claude. Rather than handing those agents your raw data, show up inside them as your own agent, on your terms and under your name. Let the generalist handle the commodity requests and keep the high-judgment work, and the context it generates, for yourself.
Conclusion
In all three cases, the pattern is the same. Engineering teams trust their tests, so the agents get the job. Consumers trust a personal agent, so it gets the results. And B2B companies that give up the experience give up the trust and the task goes with it.
We at Angular are spending time with teams figuring out which tasks they want to own. The models will keep getting better. The trust still has to be earned.
Jerry Dischler
FROM THE BLOG
Could the future of software be fluid
How do we get the best of AI without losing the soul of software?
The future belongs to young missionary teams
Why it makes more sense betting on youth in the current moment
The AI-native enterprise playbook
Ten real-time observations on a rapidly evolving playing field
No more painting by numbers
It’s the end of the “SaaS playbook.
WORTH READING
HARD TECH
Work like an ANT. Anthropic has filed to list on Nasdaq under “ANT”, and the S-1 is the first audited look inside a frontier lab. Revenue was $16.3bn in the first half of 2026, up from $1.2bn a year earlier, and net revenue retention was around 900%. But, about 20% of first-half revenue came from just two unnamed customers. The seven co-founders will keep a voting majority and there were some other relatively unusual admissions for an S-1 filing, with Anthropic continuing their narrative of the existential risk of AI. In the wake, posters on X highlighted many of the “insane admissions” from the filing about AI risks (extinction potential, self-preservation, resistance to shutdown, blackmail/deception behaviors, etc.)
Chonky. Mistral has released the full weights of Mistral Large 4, nicknamed “Le Chonk”: a trillion-parameter multimodal model that companies can download and run entirely on their own servers. Three extra weeks of reinforcement learning after its October preview lifted its Terminal-Bench coding score from 28% to 41%, which Mistral says makes it the best open-weight coding model, level with China’s GLM-5.3. Mistral is pitching it squarely at sovereignty and security, as the largest open-weight frontier model from a Western lab, with Microsoft the first cloud to carry it.
HARD MARKETS
Optics and the data centre. Marvell is buying Lumentum in an all-stock deal valuing the laser and optical-switching maker at $115bn, betting that AI clusters are hitting the limits of copper and that the future belongs to whoever co-designs the electrical and optical stack together. Lumentum CEO Michael Hurlston told staff he has long seen optics and electronics “beginning to collide”, but the market was less bullish, and Marvell’s shares closed down 12.7% on the announcement. Jensen Huang, whose NVIDIA backs both companies, has called Marvell the “next trillion-dollar company”, but Morningstar warned before terms were announced that if current results prove even partly cyclical, “Marvell could be overpaying for a weaker asset than it bargained for.”
Data in orbit. Google says its Project Suncatcher prototype has sent back the first AI response from a data-centre-class TPU in orbit, running its Gemma 4 model to write a haiku about being in space. The satellite, built with Planet and launched on 1 October aboard SpaceX’s Transporter-18 rideshare, carries four TPUs powered by 1kW of solar panels, and will now be tested for how the chips cope with radiation and thermal stress. It’s a proof of concept rather than a business, and Google admitted that space data centres currently cost around 100x more than terrestrial ones, and launch costs would need to fall dramatically before the economics work.
The fastest new nuclear is old nuclear. The US Department of Energy has offered Constellation a conditional loan of up to $5bn to uprate eight existing plants, squeezing 1.1GW of new capacity out of 15 reactors across Illinois, Maryland and Pennsylvania. Uprates deliver power in a fraction of the time and cost of new builds or restarts. The money follows the tech giants: Constellation had already committed more than $7bn for roughly 1,080MW off the back of 20-year contracts with Google and Amazon. The loan isn’t final yet, and TechCrunch notes there’s no evidence the extra capacity will lower bills for residential customers.
HOW TO STARTUP
Internal growth. Lenny Rachitsky talks to Hector Serrano, who started as a 19-year-old intern and built the growth engine behind Wispr Flow, the $2B AI voice-to-text company. As Head of Content, he runs a roster of 50 creators producing 500–600 paid user-generated content ads every month. They cover why paid UGC is one of the biggest untapped growth levers right now, the step-by-step playbook for building a programme, when to start spending on paid ads (only after product-market fit), why the best marketers are “allergic to AI”, and why YouTube is the most underrated growth lever in SaaS.
HOW TO VENTURE
Thinking outside the box. In a reflection on his conversation with BoxGroup’s David Tisch Harry Stebbings publicly recants his view that a fund backing 150+ companies with $200M+ under management couldn’t produce outsized returns. In the episode Tisch walked him through BoxGroup’s numbers, with Cursor, Baseten and Clay in the portfolio. Tisch reckons concentrated funds claiming to win 90% of the deals they want actually win around 20%, quietly offering terms to 120–150 founders to land 30, which he calls adverse selection nobody admits to. He also argues that “we don’t need pro rata” is marketing garbage because the right to follow on is earned, and BoxGroup bars partners who haven’t met the founders from being negative about a deal, since with 70–90% of companies failing, explaining why one won’t work is obvious and irrelevant.
Slow your roll. Yoni Rechtman shares the thinking Slow Ventures presented to the ~25 institutions behind roughly 90% of its capital, summed up as neutral to short on AI beta, long the alpha. Slow is backing companies that sell outcomes rather than software, the second-order effects of AI flipping the old assumption that effort costs more than attention, the bottlenecks, physical assets and data at the edges of the AI supply chain, and technical, commercial, rangy “post-AI” founders. It’s passing on infinite capital consumers where its check doesn’t matter, paying up for legibility, “software can do cool stuff” pitches with no view on capturing value, and trades dressed up as investments.
PORTFOLIO NEWS
Reco has been named to this year’s Fortune Cyber 60, Lightspeed and Fortune’s list of venture-backed cybersecurity startups.
PORTFOLIO JOBS
FalkorDB
Core Developer (Tel Aviv)
Angular Ventures is a specialist early-stage enterprise tech VC firm based in London and Tel Aviv. Angular backs companies born in Europe or Israel with the ambition to define a category and achieve global leadership, usually by starting with the US market.
/
If you are running an early-stage start-up in the enterprise space anywhere in Europe or Israel, we’d love to hear from you to see if Angular can help.
Update your email preferences or unsubscribe here
© 2026 The Angle
3234 Great Marlborough Street
London, England W1F 7JD, United Kingdom
Powered by beehiiv