Building an AI-native services company.

Would love to connect with founders, agency owners, and industry folks to learn, collaborate, and grow together.

#LetsGoX

You’re running a SaaS.

Your AI support agent promises a feature you haven’t built.

A customer renews for $20,000 because of it.

Building it delays everyone else’s roadmap by 2 months.

How would you handle the promise?

You’re hiring two developers.

One has 10 years of experience but barely uses AI.

The other has 2 years of experience and ships 3x faster with AI.

Both produce equally good code.

Who gets the offer, and why?

If Codex and Claude Code are supposedly the most powerful AI coding tools, why are people still choosing Replit, Lovable, and Emergent?

What’s your first move when someone clones your startup overnight and somehow gives it a better landing page?

me after claude finally fixes the bug it introduced 20 prompts ago:

What’s the biggest lie people tell themselves when building a startup?

What’s one thing AI made you worse at?

You’re interviewing a developer.

They finish your take-home assignment using Claude Code. Every test passes. They wrote almost none of the code.

What’s the next question you ask before deciding whether to hire them?

Before you blame the model, read the input it got.

Most "the agent is dumb" bugs I've chased were the agent being handed a stale record, a truncated doc, or last week's prompt.

Log the exact input on every run. Most mysteries end there.

Your AI API returns no answer. You still get billed.

For reasoning models, max_output_tokens covers hidden thinking AND visible text. Set it too low and the model may stop before printing a word.

Check status: incomplete, then leave room for the answer.

Your AI app feels slow. Switching models might not fix it.

Measure 2 things:

TTFT = time until the first token (queueing + prompt processing).
ITL = time between streamed tokens.

vLLM exposes both. If TTFT spikes, inspect the queue before tuning generation.

Your AI agent shouldn’t use vector search to answer “How much did we sell yesterday?”

That’s a SQL question.

DuckDB queries Parquet files directly:
SELECT SUM(amount) FROM 'sales.parquet' WHERE day = CURRENT_DATE - 1

Use embeddings for similarity; SQL for exact totals.

Two coding agents editing one folder can overwrite each other's work.

Use Git worktrees:
git worktree add ../feature -b feature

Each agent gets a separate directory and branch, sharing one repo. Edits stay isolated; review and merge them when done.

anyone here building ai agents for boring back-office work?

invoices, intake, follow-ups.

what's the one task clients actually pay for? drop it 👇 lets #connect

Giving an AI agent 80 tools upfront is like importing your entire codebase to call one function.

OpenAI Tool Search can load definitions on demand: add tool_search and mark rarely used functions defer_loading: true.

Less context wasted; one extra discovery step.

Before you add an LLM step to a workflow, try writing it as a rule first.

If the rule fits in one line, you don't need a model there. You need an if statement.

The check I run on every step:

1. Input is structured (a status, an amount, a form field): rule
2. Output is one of a few known values: rule or a lookup table
3. Input is messy (an email thread, a PDF, a phone photo of an invoice): model
4. A wrong answer costs money or trust: model drafts, a person approves

Rules are cheap per run, fast, and fail the same way every time, so you can actually test them.
Model steps are where the cost, the latency and the weird 3am failures live. Spend them only where something needs reading or judgment.

A good agent is mostly boring code with a few smart steps in it. Not a model in a loop deciding everything.

Google monetized what you search.

Meta monetized what you scroll.

Personal AI bots could monetize what you’re about to do.

Imagine brands bidding to become your AI’s recommendation.

The next trillion-dollar ad industry might not sell impressions.

It might sell decisions.

Git should require a doctor's note for commits made after 2 a.m.

"I know what I'm doing" is not a valid medical diagnosis.

I'm completely rebuilding Vyra's(agentic gtm) multi-agent architecture.

No supervisors.
No fixed agent roles.
No endless agent-to-agent handoffs.

I'm going all-in on ScaleMAS: agents coordinating through a shared knowledge substrate.

Time to see if the research holds up in the real world.

here is the paper: arxiv.org/abs/2610.16120