Krishna

There's a little too much "build only for agents" hype right now. Obviously, there's loads to build for agents. But if you are actually using agents, you know they need to be managed. So, building for humans who manage agents is a totally viable play.

@jasonlk Ya this can get out of sync so quickly. You build / change something but forget to update the context for your AI agent. We try to do this by just having Claude Code do a sweep regularly.

The other day the AI Agent for a leading AI startup told us we had to pay 4x more than we had been paying to keep using it It tuns out we didn't have to -- but the AI Agent hadn't been properly trained on its own product You have to, have to, have to train all your own AI Agents too Or out comes garbage

@adisingh Still remember the day you reached out. Incredible to see you folks have the conviction and keep building. Well deserved :)

POV: You didn't give up. For most of AgentMail's life, growth was flat. No one understood it but we kept pushing.. until one person did. The next 10 followed. Then, the rest. Keep going, make that bet on your vision for the future. Make something agents (should) want.

@ivanburazin At this point, I can’t think of why you would hamstring an agent by giving it just a code execution box. Maybe if the scope was super limited or something.

Even 6 months ago, the market was generally skeptical of sandboxes as a primitive. For most workloads, you could just use code exec boxes as they're cheap and fast. But they were designed for code that behaved predictably with expected inputs/outputs, and were stateless by default. When autonomous agents arrived, they needed environments to write, preview, and run code + full computer access to browse and execute tasks. Each task required a persistent, isolated environment that could spin up in milliseconds and run for hours. Now multiply that by thousands of parallel agents serving millions of daily active users. With compute becoming infinitely forkable, sandboxes are now the substrate for the agentic economy.

Question for people building API products: how worried are you about an AI model just doing that natively with time? A good example of this is structured data extraction

@ay_ushr Huge need for the infra so big opportunity if you can win it

i think at this point there's more ppl building agent sandboxes than actual agents

@thorstenball I don’t think all of it goes away. Some of these markdown files are general and others are specific to a company (like an SOP). The latter stays

Lately, whenever I open this app and see the latest tricks, and hacks, and notes, and workflows, and spec here and skill there, I can't help but think: All of this will be washed away by the models. Every Markdown file that's precious to you right now will be gone.

@amrishrau Slightly different perspective. PMs can now become executors, meaning actually get the work done themselves. More that technical, I think becoming great at reviewing AI’s output and systems (like evals) to improve the AI system will be extremely valuable.

Product Management will see a massive change. PMs will become lot more technical or they will become product marketing managers. The current won’t exist.

@TaraViswanathan I think this intentional as they are going after businesses?

This won’t be OpenClaw until they give it a personality. I don’t feel a connection to “Cowork”. But I am obsessed with Kevin, my OpenClaw minion, who runs around doing all kinds of things for me while saying hilarious things in group chats with friends and making me laugh pretty much every day.

@petergostev Exhausting. Reviewing AI output to improve quality is absolutely required but also exhausting

There's worry that people will stop using their brains with LLMs, but managing several AI agent threads in parallel has been some of the most cognitively intensive work I've done in years

@pestctrlguy Depends on whether you’ve finished building your thing! I think the bigger cost is always going to be maintaining the internally built CRM

Question: What happens when PE, VC, and public companies stop subsidizing AI and all of these people that built their own CRMs and tools can’t afford their own setup anymore? One of these days these companies will charge an appropriate amount and it’s going to bust budgets right?

@boringmarketer Would love to have you use our product @with_ako for your service if you’re open to it. I’m using it internally for our own GTM so I can share why it works well

i'm going to take on one pilot customer for a new managed ai "go to market" service yes AI, agents, and automation is deeply involved, and myself/team are the humans in the loop who will turn the dials and interface with your team ideal fit is a b2b company (software or service) that is doing at least $2m/year, you want to leverage ai for marketing and you know you could be moving faster at a higher quality hit me with a dm if this sounds interesting and I can share more details

@thdxr We did some preliminary code mode evals in a JS sandbox when we wrote our SQL vs. Bash eval, and JS was about 2.5x more token efficient than bash to achieve the same result.

@hrishioa Flash is more expensive than Haiku? Don’t think so

You're all sleeping on Haiku. Haiku 4.5 is a new beast - it isn't the 'glorified autocomplete' we've come to think of it as - it's a capable, strong agentic model that's crazy fast and dirt cheap. Here's Haiku doing deep research with exa, generating reports, making infographics, integrating those infographics into the report, all around five cents, in under two minutes. Flash is a strong contender but it's more expensive, slower, cannot do interleaved toolcalls - but it does have the 1m context. Once you have a good harness - like this internal chat agent we use, or hankweave - the model matters less than it did before. You only need Opus when you're surrounded by twenty confusing skills that turn on at random times, with confusing compressed context. Use haiku more - it'll highlight some fixable problems. Most hanks (AI Programs) on hankweave move towards sonnet from opus and then to haiku as they mature - which means faster and cheaper executions overall, and more breathing room for experimentation.

@rjs Our product does this natively and operates over Slack / email etc. Context builds across the whole team and it’s shared. Can do a load of tasks for you. Let me know if you’re open to trying: @with_ako

Starting using a shared memory git repo with team members who are Clauded. Now when I want to explain something to them, I am oddly inclined to give them a prompt. "Ask this." Interesting tooling gap.

@cjc Really does remind of blockbuster choosing to ignore Netflix kinda thing..

@antinertia I have a feeling AI changes these numbers dramatically. Channels get flooded, CAC becomes very high etc

you all see companies going from $0 → $100m in months… but here’s the real math you should actually care about for plg saas / prosumer: - ltv:cac = 3:1 to 5:1 lower = not efficient higher = underinvesting - cac payback < 6–9 months how fast you get your money back after acquiring a user - nrr > 100% how much your existing customers grow your revenue (after churn) these are a few of the key metrics you need to track as a founder, and what your investors will ask for others include: - expansion revenue drivers what makes users pay more over time (credits top-ups, etc.) - one-and-done rate % of users who use once and never come back (this is very high in ai right now…) this is different for sales-led products (e.g. enterprise) just wanted to write this because twitter feels like an echo chamber, and no one talks about the real math behind those tech companies

@ivanburazin 100%. Think there’s a big diff between standards / processes (slack to communicate) and CRUD interfaces. Embedding into existing processes helps with adoption

Agents won't vibe-code a new Slack every time they need to communicate. They'll use the same Slack because the other agent's team also uses Slack. Traditional SaaS will survive because standards don't get disrupted by probabilistic code generation. Network effects and standardization still matter.

@jessethanley Good way to put it. Do you feel this is particularly acute given the use case Bento serves?

I try to avoid being a doomer, but I do feel immense pressure to have my product conform to the new way of working in a way I haven’t felt before. Organic traffic is no longer a distribution strategy, customers don’t want to login anymore, you need to have a CLI/MCP or customers will go elsewhere, our AI features need to be updated weekly and one-shot the most random requests… Your company is at risk if you don’t move fast enough. I’m embracing it but I understand why people are anxious. It’s never ending.