Never present an invoice via email; always over the phone
The host states a rule of sales: never present an invoice via email. Present it over the phone so you can answer questions, resolve concerns, and get them to buy.
35,848 results · Most relevant
The host states a rule of sales: never present an invoice via email. Present it over the phone so you can answer questions, resolve concerns, and get them to buy.
If you say your super secret env var can only be used for the SEC and Claude tries to use it for anthropic.com, the swap won't happen — but when it tries to call sec.com, the swap happens for that. For example, you could install a Google Workspace CLI on your sandbox and give your API key as the environment credential — vaults will follow the same placeholder-swap method.
The answer to the ontology question is that there is no universal ontology — there never was one. The good news is that this makes every ontology and schema valid, as long as you can label and preserve the context around it.
The fourth and most important thing missing is a failure surface. The failure surface has been the hardest thing to put into words and into tooling and instructions that help models get there; the current best models are still only barely good at this — sparks of intelligence, so to speak.
The team rewrote the questions to define what counts as enough evidence, which fixed model failures without changing models or thresholds.
In a system like this, ontologies exist as intermediates between intelligence and result, which makes them easier to reuse. A just-in-time ontology is built when the job needs it, kept with its receipts, and reused when the next job needs the same thing. Once you have just-in-time ontologies, you can make these things cheaper and run them in parallel across data sets to scale up.
Third-party outbound tools are limited in the amount of first-party context and data they can bring in — all of them. There's no way a third-party outbound prospecting agent could know all their proprietary context. The ability to iterate on the response is the other biggest limitation. These tools are great at the initial — opening the conversation, getting a meeting — but terrible at follow-up because they don't know what to do; once you have an interaction with the person, all of these apps kind of fall apart.
Described principles from the host's experience; generalization across 'all' tools is his claim.
The in-person (IRL) world has become a critical part of the go-to-market playbook for agentic players, which was not clear 18 to 24 months ago. OpenAI and Anthropic are brutally competitive in the enterprise, driving them to run their own events and attend third-party events — OpenAI just did their own dev day. They are signaling to the market that they need their own and third-party events to compete and get buyers.
Interpretation of hiring and event signals; the competitive framing is the speaker's read.
Enterprise buyers want to talk to vendors, qualify them, compare them to competitors, and do it efficiently at events. As deals get bigger and buyers become enterprise buyers, they don't test products on their own; they want to talk to people, take notes, evaluate vendor criteria, and get questions answered their way. For a buyer wanting to meet six vendors at once, going to one of the leading industry events is very efficient. Meeting many vendors in one place is more important in this era than ever because of the explosion of vendors and how much has changed. Buyers with six-figure budgets want to figure out what decisions to make in person and cannot do it all on Zoom.
Consequential creator advice requires independent verification before Today can promote it.
Vendors that materially charge for agent API access risk agents routing around their systems of record entirely. An agent that can only access a mission-critical API for 10–20 minutes a day will immediately route around that vendor, and agents will not recommend systems of record that materially charge for API access.
Cost control starts with model selection. Each LLM step in an Orbit workflow declares its own model, so a planning step and an implementation step can use different models; matching the model to the task keeps cost under control.
Matt Ragone argues that the people doing the work should also be the ones building the skills: the person who knows insurance best builds the insurance skill, and the person who best knows tax code and tax law should build the tax one. When subject matter experts build for their own niche, the skills get a lot more specific.
Matt Ragone says two things had to be true for agents to work: the agent has to work where the work lives rather than in a separate chat window you copy and paste into, and it has to know how your firm does it — not how financial planning works in general, because the model already knows the general part. Firm-specific details include what to look at first, what to always flag, and how to phrase a specific recommendation.
When teaching an agent, the narration of judgment is what matters: not just 'next I open the pay stub' but why you're opening it, what you're looking for, and what would make you stop and ask a question. The steps are easy to write down; the judgment is what makes it your firm's work instead of generic advice.
The architecture provides human-in-the-loop with access control that's actually backed by policy; access is controlled at the task level by the security token service and policy, while retaining all key capabilities.
Forcing human-in-the-loop approval just passes access control on to a consent-fatigued user, shifting responsibility up the chain even though we've spent decades solving access management for humans. You also can't apply it everywhere because humans make mistakes and many agents run autonomously.
Agent incidents happen because humans give agents way more access than they need, and overprivileged agents can act freely on their own decisions even when they have supervision.
With this architecture, agent access problems are solved: the user and agent identity are known all the way down the chain, and the agent has task-scoped ephemeral access to the MCP server and downstream resources.
Simply removing one or more lethal trifecta capabilities makes the agent significantly less useful, so the goal is to keep all three while staying safe: control the agent's access so the attack surface is minimized and the blast radius is contained while preserving capability and autonomy.
The tempo fingerprint over NVLink achieves 79.2% model identification accuracy against a 20% baseline across 5 models. The layer fit classifier also picks up fine-tunes and distillations sharing the same base model, and works on proprietary or unreleased models it has never seen before, since architecture and layer count carry over unchanged.
Choose sources → to get a library of your own and a daily brief.