Principle · MoneyNikhil Harithas describes Factory's key differentiator as building a custom-fit deployment of the product in the customer's environment on the customer's terms — respecting their security constraints, their compute authority choice, and running totally in their VPC. He reports this approach has seen significant success because it builds trust in the customer relationship.
AI Engineer · Your AI Agents Shouldn't Care Where They Run - Nikhil Harithas, Factory
Report · MoneyIn the Open Compute architecture, a client sends a request to the backend, which routes to a compute gateway that invokes an Open Compute driver to spin up compute (e.g., an EC2 machine) in the customer's VPC. That compute reaches out to Factory's relay in the control plane and keeps an open WebSocket connection; requests from the CLI, Slack, or web app go to the relay, which relays them to the sandbox. The LLM call can be established as a bring-your-own-key (BYOK) situation — e.g., via Bedrock — so even inference stays within the customer's VPC. Nikhil notes BYOK is only relevant for companies that already have a deal with one of the labs.
AI Engineer · Your AI Agents Shouldn't Care Where They Run - Nikhil Harithas, Factory
Creator perspectiveJacob Lauritzen describes the old approach on a big case with a million documents: throwing a team of paralegals locked in a room sitting there until 4am. That doesn't need to happen anymore because extraction and reprompting can be done instead — he no longer needs a paralegal for tabular review of a million documents.
hedged
Latent Space · Building Legal Agents When You Can't See the Data - Jacob Lauritzen, CTO of Legora
PrincipleLegora now post-trains models only for specific bounded tasks where there is a huge delta between what a frontier model can do at cost and latency versus what a post-trained model can do. Legal research is an example: it is a fan-out retrieval problem where a purpose-built model can run queries super quickly, super cheaply, and at low latency.
Latent Space · Building Legal Agents When You Can't See the Data - Jacob Lauritzen, CTO of Legora
Creator perspective · MoneyDharmesh Shah's framing: treat chat interfaces like ChatGPT and Claude Cowork as the new operating system / new browser — the consumption layer for value — and build products that plug into them, as SaaS companies did when the browser displaced client-server. To evaluate whether an AI interface erodes your product's value, place your business on an interface-value spectrum: some products' value is entirely the interface, others' value is the network or work behind it.
The a16z Show · HubSpot’s Co-Founder on Finding Your Next Startup Idea
Principle · MoneyDetect secrets on user prompts and tool requests, and block or quarantine them before the model sees them. Once interception happens, log it so there's an audit of the event — inspected, blocked, and logged; the secret never left the machine.
AI Engineer · A Valid Key Let an AI Agent Delete Production - Sagar Batchu, Speakeasy
Creator perspective · MoneyThe flag agent scores every flag the way an experienced reviewer would: how many modules does it touch, is it multivariant, does it share a component — then it checks the experiment data: is the rollout frozen, is there a sample ratio mismatch, is there a variant already rolled out to 100%? Only the safe mechanical cleanups ever reach the model.
AI Engineer · Scale the Judgment, Not the Model — Andrew Orobator, Reddit
Creator perspective · MoneyPapers from Princeton and cloud-based robotics show that designing a harness where robot-facing tools look like computer-use tools — an agent dragging a cursor to control where the robot goes — improves LLM performance on physical tasks.
Y Combinator · Robot-Use Agents: Why General-Purpose Models May Win in Robotics
Creator perspective · MoneyMost people use LLMs via in-context learning by appending examples to the context. ICL improves validation performance greatly and very cheaply — no SGD, no FLOPs cost — and is remarkably effective in low-data regimes. But improvement is non-monotonic (worse, then better, then worse, then better, aggressively), and ICL caps out very quickly: after about 20, 30, maybe 40 examples, improvement is basically saturated and more examples do not improve performance.
about 20, 30, maybe 40 examples; remarkably effective
Y Combinator · Robot-Use Agents: Why General-Purpose Models May Win in Robotics
Report · MoneyThe source endorses asking your champion to kick off procurement, security, and legal at the same time instead of one after the other the moment you are picked as vendor of choice — a move that alone can turn a six-week review into two weeks.
30 Minutes to President's Club | No-Nonsense Sales · #625: 18 Deal Cheat Codes to Cut Your Sales Cycle in Half
Report · MoneyThe host endorses using a redline deadline — a set of deadlines for when the customer's legal team delivers their first round of redlines, when the seller returns a response, and when the customer delivers their second round — calling it one of the most effective tools for project managing customers through the legal process. Setting it incentivizes the customer to hit those dates even without a give-get attached, and gives the champion a way to enforce moving through the review queue for their legal team.
30 Minutes to President's Club | No-Nonsense Sales · #625: 18 Deal Cheat Codes to Cut Your Sales Cycle in Half
Report · MoneyThe host endorses pre-scheduling the kickoff call before the contract is signed (deal cheat code #17). Pretty much everyone will say sure, let's go do that; a signed contract is required to hold the slot; and the pre-scheduled call gives the champion a reason to hold their legal and procurement teams accountable to the agreed timeline. The kickoff call covers data migration, integrations, and related implementation topics with the implementation team.
30 Minutes to President's Club | No-Nonsense Sales · #625: 18 Deal Cheat Codes to Cut Your Sales Cycle in Half
Report · MoneyAri Weinstein describes OpenAI's enterprise controls (part of Astra) that let businesses choose how much human-in-the-loop approval the agent requires, including a Guardian system where a smart model auto-reviews every agent action for safety and alignment with original intent.
The a16z Show · OpenAI: When AI Agents Use Computers Faster Than You
Report · MoneyJames Sun describes enterprises as a natural fit for computer-use agents because they run legacy software around 15-20 years old with no MCP tools and confusing UIs that are hard for normal people to operate, and notes companies are building skills that teach the models to use that software better.
The a16z Show · OpenAI: When AI Agents Use Computers Faster Than You
Report · MoneyJames Sun describes Astra as OpenAI's most aligned model, much better at resisting prompt injection and avoiding unintended actions, which he says makes enterprises feel safer using it for sensitive work like procurement or vendor communication.
The a16z Show · OpenAI: When AI Agents Use Computers Faster Than You
Creator perspective · MoneyAri Weinstein sees promising enterprise automation use cases in two big categories, starting with automating tedious workflows: ERPs, CRMs, and every system of record; finance, where many financial software programs lack APIs for quickly pulling data and teams do manual month-end reconciliation across multiple ledgers; procurement, where teams fill out giant spreadsheets that ops manually enter into internal software before the procurement team contacts vendors; and sales, where system-of-record data entry is the tedious part agents can automate. Within company-approved software, he notes, automation is straightforward because companies want people to automate tedious work.
hedged
The a16z Show · OpenAI: When AI Agents Use Computers Faster Than You
Creator perspective · MoneyDaksh Gupta describes how Greptile disintermediates all context sources with a living wiki that acts as a world model of how the company works, incorporating codebase knowledge, production telemetry and logs, and roadmap information from documentation and Linear tickets.
AI Engineer · How to Build an AI Code Review Agent - Daksh Gupta, Greptile
Report · MoneyKevin Hou announces that agent teams are in public preview inside Antigravity: typing the slash command '/teamwork' enters a new mode that unleashes a swarm of agents onto the task at hand.
AI Engineer · Get Out of the Model's Way — Kevin Hou, Google Antigravity
Report · MoneyKevin Hou reports that researchers were able to automate 90% of the side-by-side eval workflow by simply asking the agent about the evals in question using natural language; all the user had to do was load up a skills file and ask away.
AI Engineer · Get Out of the Model's Way — Kevin Hou, Google Antigravity
Creator perspective · MoneyRoland Gavrilescu lays out what he says you need to have a frontier agent: codified evals, a way to run evals, loops to continuously improve those evals, processed signals, the right tools for certain models, and different harness profiles for different models.
AI Engineer · The Loop Is the Product — Roland Gavrilescu, Introspection