Recommendation · MoneyThe source recommends starting with just one grow kit block on your kitchen table as the bare minimum space needed to get started. Once comfortable, make your own little grow hut; there are good kits online that have the whole system already built in. With a couple tiers of a grow hut, you can grow lion's mane and oyster mushrooms, harvest them, and sell them or eat them.
hedged
UpFlip · Here's How to Start a $10K/Month Mushroom Business from Home
Recommendation · MoneyRecord each agent run as a chain of steps, where each step has an ID, inputs, and outputs, so any point along the run can be interrogated later to understand what was done at each step.
AI Engineer · Your Financial Agent Needs to Show Its Work - Anirudh Veeraragavan, Fieldguide
Recommendation · MoneyAgents should be allowed to draft unreviewed rows as themselves, but must never overwrite what a human reviewed. Agent output exists as a suggestion, not authoritative truth.
Also discussed by Orca guest on Latent Space (unclassified source — not counted until classified). Not counted as independent sources.
AI Engineer · Your Financial Agent Needs to Show Its Work - Anirudh Veeraragavan, Fieldguide
Recommendation · MoneyMaintain a record of the steps that happened in an agent run and the culmination of the run, so each point along the way can be interrogated. Structure an agent run as a chain of recorded steps where each step has an ID that lets you understand what is being done, the inputs, and the outputs. Clearly delineate between deterministic steps where code fetches an item or does a match versus steps where the agent extracts a field from a PDF or arrives at the final conclusion.
AI Engineer · Your Financial Agent Needs to Show Its Work - Anirudh Veeraragavan, Fieldguide
Recommendation · MoneyTreat a citation as identity, not just a string: a very specific ID, a very specific version of the document, and a hash of the contents behind it. Concretely tying where an agent picks something from allows you to assert rules around the evidence that's allowed and any quality constraints you want around it — for example, only allow evidence sourced directly from the bank, as opposed to something the client uploaded, where they may be able to tamper with it.
AI Engineer · Your Financial Agent Needs to Show Its Work - Anirudh Veeraragavan, Fieldguide
Recommendation · MoneyThe agent should stop and ask a human when it is trying to change work a human has already reviewed, when a new run is disagreeing with a previous run, and when evidence is not conforming with the established rules — all examples where something has gone awry and a human needs to step in to determine what's happening.
AI Engineer · Your Financial Agent Needs to Show Its Work - Anirudh Veeraragavan, Fieldguide
Recommendation · MoneyTo make an agent followable, put controls in place around authorization, execution control, evidence provenance, and evaluation integrity. Defensibility is an AND: you need all four, otherwise things can always slip through the cracks. All agent controls ultimately rest upon knowing when the agent should stop and the human should decide instead.
AI Engineer · Your Financial Agent Needs to Show Its Work - Anirudh Veeraragavan, Fieldguide
Recommendation · MoneyEscalations to humans should be easy to verify but not necessarily easy to approve. Automation bias is real: if you continually ping the human, they'll just start hitting yes to everything. Only flag exceptional cases for human escalation, and those exceptional cases should very clearly call out what is wrong and where someone needs to step in.
AI Engineer · Your Financial Agent Needs to Show Its Work - Anirudh Veeraragavan, Fieldguide
Recommendation · MoneyA citation is an identity, not just a string: it should have a specific document ID, a specific version of the document, and a hash of the contents behind it.
AI Engineer · Your Financial Agent Needs to Show Its Work - Anirudh Veeraragavan, Fieldguide
Recommendation · MoneyAudit your eval/gold dataset as well, because a reviewed audit is not necessarily a correct one. Mistakes in the dataset can punish a right agent or reward a wrong one.
AI Engineer · Your Financial Agent Needs to Show Its Work - Anirudh Veeraragavan, Fieldguide
Recommendation · MoneyOnly flag exceptional cases for human review rather than pinging the human continually. Each escalation should clearly call out what is wrong and where someone needs to step in, and humans stepping in for the flagged cases should have the context they need to arrive at the right decision.
AI Engineer · Your Financial Agent Needs to Show Its Work - Anirudh Veeraragavan, Fieldguide
Recommendation · MoneyMake sure agent traces contain all the information up front, not just transcripts. Include everything connecting the agent to the outside world — tool calls, file system calls, and MCPs — plus the outputs, such as code diffs living inside conversations or the actual pull request if you pin the commit and look at the diff. In short: the transcript of what was said, the tool calls showing what the agent did, and the output you may want to grade for quality.
AI Engineer · Your AI Agent Is Hiding Its Mistakes From You - Ben Holmes, Warp
Recommendation · MoneySet up scoring agents to score the quality of how agent conversations went — think of it like an LLM as a judge reviewing against rubrics that you define, across metrics like efficiency, code quality, or test coverage. Keep the rubric as succinct as possible and build it up as you go; you can set up a score with literally just a couple of bullet points and your rubric and test it out yourself. Set labels that define what success and failure look like on the scoring criteria, and you can decide what pass and fail look like for your scores.
AI Engineer · Your AI Agent Is Hiding Its Mistakes From You - Ben Holmes, Warp
Recommendation · MoneySet up a scorer called redundant test that evaluates how often the agent writes too many tests across each conversation. Set up a verbosity metric scoring whether the agent is overly verbose when talking to you or writing code — how well the agent communicates changes matters because poorly summarized changes and excessive comments can be barriers to getting a code change reviewed quickly. Set up an efficiency score that looks through tool calls and asks whether the agent efficiently got to the solution or spun its wheels, i.e. worked toward the result or wasted effort on redundant steps and unnecessary back and forth.
hedged
AI Engineer · Your AI Agent Is Hiding Its Mistakes From You - Ben Holmes, Warp
Recommendation · MoneySalesforce follows a disciplined internal M&A framework with three elements: (1) strategic fit and how much customer success the acquisition can drive, (2) whether it can accelerate the business, fit the technology stack, and align culturally with the team, and (3) how to deliver value to shareholders and customers, the timeline for value accretion, and whether the valuation is reasonable. Somorjai notes the valuation-reasonableness element is the hardest right now because everything is expensive.
hedged
The a16z Show · Inside Salesforce’s $100 Billion M&A Playbook | John Somorjai
Recommendation · MoneySomorjai advises founders preparing for an acquisition to be ready and organized so data flows during diligence, since many companies lack strong controls over their financials, lack data on customer contracts, and don't understand what customer attrition really is — and ARR is a wonderful metric that masks attrition. He says it's important to understand the attrition of your customers and do something about it, because very high attrition usually means a lot of unhappy customers and possibly product quality issues, and high customer attrition is one of the big red flags Salesforce looks for. He also advises red-teaming your technology stack before diligence: the acquirer will do its own penetration testing and find the problems, and you definitely want to identify the vulnerabilities they
hedged
The a16z Show · Inside Salesforce’s $100 Billion M&A Playbook | John Somorjai
Recommendation · MoneyPhilipp Schmid recommends placing a network proxy around the sandbox that watches all traffic in and out. The model gets a local dummy value for credential environment variables, and when a call matches the user-defined domain, the proxy swaps the dummy value with the real credential, so the agent never knows its dummy token is not working.
AI Engineer · 7 Things Your AI Coworker Needs - Philipp Schmid, Google DeepMind
Method · MoneyPhilipp Schmid explains how to use managed agents: to use a managed agent, instead of providing the model ID, you provide the agent ID (e.g., AntigravityPreview09.com), give it a prompt, set the environment to remote so it runs in the managed sandbox, and that's all it takes. To define your own agent, you use a simple API: give it an ID (the name you'll use later to call the agent), set the base agent (the underlying agent harness), define the model, and set a system instruction. The remote agent uses the same agent harness as Antigravity, so if you use Antigravity as your coding agent, you get the same agent harness remotely. To call your custom agent, use the Interactions API create call and instead of the model ID for Gemini or the default agent, use your own agent ID (e.g., 'bug').
Method, not a verified protocol.
AI Engineer · 7 Things Your AI Coworker Needs - Philipp Schmid, Google DeepMind
Recommendation · MoneyDr. Pedneker recommends that the right-problem consideration should lead to redesigning the workflow, or at least re-evaluating what to focus on. He notes this question of what to focus on applies across research capacity, decision speed, data quality, latency, and reliability — not just finance.
AI Engineer · Your AI Isn't the Advantage. Your Judgment Is - Sid Pednekar
Recommendation · MoneyDr. Pedneker recommends thinking about AI advantage in terms of three verticals: design the right problem, build shared capabilities, and compound the advantage. He argues this framework applies to organizations and individuals alike, and to functions regardless of their position in the organization.
hedged
AI Engineer · Your AI Isn't the Advantage. Your Judgment Is - Sid Pednekar