GoZen Technologies, Morris Plains, NJ (2026)

08/06/2026

Most teams ship AI agents the same way: try it a few times, it looks good, push to production, hope.

Then a customer asks something you never tested, the agent gives a confident wrong answer, and you find out from an escalation.

The fix isn't a better prompt. It's evals.

An eval is three things: an input, the agent's output, and a scorer that says how good that output was. Run it across a set of real examples and, for the first time, "good" has a number instead of a feeling.

Using a support agent as the example, here's what eval engineering actually looks like:

Score with the right tool. Code for the checkable facts (did the reply include the booking link?), a model to judge tone and accuracy at scale, and a human for the tricky edge cases.

Score the answer and the path. A reply that promises a refund the agent wasn't allowed to give looks perfect on the surface. You catch it by scoring the steps, not just the output.

Make it a loop, not a one-time test. Collect real tickets, score, find the biggest failure bucket, fix it, and keep the set as a regression net so the next change can't quietly break what worked.

And avoid the traps: five happy-path examples, synthetic tickets that don't look real, or a model grading a model with no human anywhere. A green dashboard that checks the wrong thing is worse than no dashboard.

Stop shipping on vibes. Ship on evals.

How are you measuring your agents today?

08/04/2026

Prompt engineering. Context engineering. Harness engineering. Loop engineering. Graph engineering.

Same field, new name every few months. The latest one is worth understanding anyway, because the thing underneath it is real.

Loop engineering is how you run one agent: a cycle of plan, act, observe, and retry until it's done. Simple, and the model drives.

Graph engineering is how you connect many: specialized nodes, explicit paths, branches, handoffs, and parallel work you design up front.

The honest version nobody puts on a slide: a loop is just a simple graph, and every graph is made of loops. You don't leave loops behind when you build a graph. You wire a bunch of them together.

So the real question isn't loops or graphs. It's whether this task fits in one loop, or has outgrown one. Answer that, and the architecture picks itself.

Start with the loop. Graduate to a graph when one loop can't keep up.

Full breakdown in the graphic.

Are you still in one loop, or have you wired a graph?

07/30/2026

The cheapest token is the one you do not process twice.

Most teams building on LLMs pay full price to re-send the same context on every call: the system prompt, the tools, the documents, the entire chat history. Only the latest message is actually new.

Prompt caching lets you reuse that stable context instead of reprocessing it. On Claude, cache reads cost about 10% of the normal input price. That is a 90% cut on the part of your prompt that never changes.

The catch is that you have to earn it. The first call writes to the cache at a small premium, so caching pays off once you reuse the same context two or three times. Low-traffic, one-off calls can actually cost more.

Three things that make or break it:

Order: the cache matches from the front, so stable content goes first and the changing part goes last.
Discipline: a timestamp or a user name in the cached prefix breaks the match every time.
Reuse: the default cache lasts 5 minutes, with a 1-hour option for longer gaps.

Agents get the biggest win here, since they loop over the same tools and history on every step.

Full breakdown in the carousel.

What is your cache hit rate?

07/28/2026

Peter Yang just open-sourced a tool that reads your writing and points out where it sounds like AI.

It's called /no-ai-slop, it runs inside tools like Claude, and it flags more than 20 low-effort patterns without rewriting a word. You see the tells and decide what to cut. It hit 1,000 GitHub stars in a day, which says a lot about how many people are tired of reading the same AI-shaped sentences.

A few of the patterns it catches:

Binary contrasts: "It's not X. It's Y."
Throat-clearing openers: "Here's what nobody tells you."
Fake-profound endings: "The future isn't coming. It's already here."

They show up everywhere once you know the shape of them, including in plenty of my own drafts.

The part worth underlining is his advice for using it well: write your first draft yourself, use AI to edit for grammar and clarity, then do one more pass by hand. Your thinking and your voice stay in the writing, and the AI just tightens it up.

I pulled the patterns, the workflow, and the context into one graphic so you can save it.

Want the tool? Comment SLOP and I'll DM it to you.

07/23/2026

AI isn't going to replace SEOs.

But it is going to replace the 3-hour spreadsheet, the quarterly audit nobody has time for, and the outreach list you gave up on halfway through.

And it will help you build the lean SEO team.

The trick is knowing the line. Agents are better than humans at anything repetitive, data-heavy, and constant. You are better at strategy, voice, and taste. So hand over the grind and keep the judgment.

Here are 5 SEO workflows an agent already runs better than a person:

1. Technical audits. It checks every URL on every deploy, not a sample once a quarter. Issues get flagged in minutes, not weeks.

2. Keyword research at scale. Thousands of terms pulled, clustered, and tagged by intent in one run, instead of a few dozen in a sheet.

3. Content briefs. It reads the whole first page, not the top 2 results, and pulls the entities, questions, and gaps you need to beat them.

4. Backlink prospecting. It finds relevant sites, gets the contact, personalizes each email, and sends at volume without running out of steam.

5. Rank and competitor monitoring. It watches every keyword and every rival around the clock and pings you the moment something moves.

Notice the pattern: speed, scale, and consistency. None of it is creative. All of it is the work that keeps you from being creative.

Let the agent grind. You go do the strategy.

Which of these is still eating your week by hand?

07/21/2026

Most outbound isn't lost to bad emails. It's lost to timing.

The lead who filled in your form at 9 am and heard back at 4 pm.
The deal that went quiet three weeks ago and nobody went back to.
The no-show that never got rebooked.
The call your rep walked into cold.
The follow-up that was going out "tomorrow."

None of that is a copywriting problem. It's a capacity problem, and it's exactly where AI agents earn their keep.

So if you're automating outbound, don't start with the clever stuff. Start with the five plays that leak revenue every single week:

01. Speed to lead. Enrich, score, and reply with a booking link in minutes, not hours.
02. Silent deals. A deal goes 7 days with no activity; the agent re-reads the thread and re-engages.
03. No-show recovery. A friendly rebook email within 30 minutes, not next week.
04. The pre-call dossier. Role, company, recent news, latest posts, ready before the call.
05. After-call follow-up. Transcript to summary, CRM updated, recap email drafted.

Every one of these is repetitive, time-sensitive, and a human adds nothing to it but delay. That's the whole test.

The plays that win aren't the clever ones. They're the ones that happen every single time.

Which of the five leaks the most for your team?

07/16/2026

AI doesn't sound less confident when it's wrong. That's the whole problem.

A made-up statistic reads exactly like a real one. A broken source link looks just as tidy. So most people do one of two things: trust the agent completely, or redo all its work by hand. Both defeat the point.

You don't need to be technical to catch mistakes. You need a habit.

Here's the one I use. It takes about 60 seconds. Call it SCAN:

S. Source it. Ask "where did this come from?" and make the agent point to the doc, link, or record. No source, no trust.

C. Check the facts. Don't audit everything. Spot-check 2 or 3 specifics (a number, a name, a date) against the real source.

A. Answer the brief. Go back to what you asked. A confident answer to the wrong question is still wrong.

N. Nose for nonsense. When something feels too neat or slightly off, slow down. That instinct is usually right.

And match the effort to the stakes: a glance for an internal draft, the full SCAN for anything a customer sees.

Don't blindly trust it. Don't redo it all. SCAN it.

Which of these do you already do without thinking?

07/15/2026

Most "AI SDRs" are just copilots with better marketing.

A copilot drafts. You still hit send. It makes one rep faster, but it does not give you another rep, so your pipeline still moves at human speed.

An autopilot runs the whole play: it finds the leads, researches them, writes, sends, follows up, and books the meeting.

Here is the honest scorecard:

Speed: autopilot. First touch in minutes, not days.
Volume: autopilot. 5,000 leads as easily as 50.
Personalization at scale: autopilot. Every lead researched, every time.
Inbound response: autopilot. It replies at 2am on a Sunday.
Control and oversight: copilot. You approve every word that ships.
Complex, nuanced deals: copilot. A human reads the room.

The volume math settles it. Ask a rep with a copilot for 50 prospects this week, and you will get them. Ask for 5,000 this quarter and the math breaks. An autopilot does not notice the difference.

But autopilot is not the answer to everything. Six-figure deals, sensitive accounts, and legally loaded wording still need a human hand.

So don't pick a camp. Set the dial: let the AI run the volume, and keep a human gate where the stakes are high.

Your reps aren't slow. They're outnumbered.

Which lane would you put on autopilot first?

07/07/2026

Most people blame the AI when it gives them a useless answer.

Usually, the problem was the brief.

AI agents do exactly what you ask, and fast. So a vague ask gets confidently finished the wrong way, and you burn an hour fixing it.

The fix takes 30 seconds. Before you hit go, tell the agent these 5 things:

01. Goal: what you actually want to happen, not just the task
02. Context: the background it can't guess (files, facts, history)
03. Boundaries: what's off-limits and what needs your sign-off
04. Examples: one sample you'd be happy with
05. Done: what "finished" actually looks like

Give it those five, and it stops guessing and starts delivering.

Swipe through for what to say for each one, with real phrasings you can copy.

Which one do you skip most often? For me, it's the example.

07/02/2026

Prompt engineering is the most overrated skill in AI right now.

Not because prompts don't matter, but because they're the easy 20%. The hard, valuable part is everything around the prompt.

That's context engineering: designing what the model can actually see when it answers. As models got better at following instructions, this is where the real leverage moved.

I broke it into the 5 layers that separate a demo from production 👇

01 · Instructions - the agent's role, rules, and guardrails
02 · Knowledge - grounding it in your data (RAG / retrieval)
03 · Memory - what it remembers across turns
04 · Tools - what it can actually do
05 · Structure - how the context is organized

A clever prompt helps once. Engineered context compounds across every query, user, and task.

Swipe through and tell me: which layer does your team get wrong most often?

GoZen Technologies

08/06/2026

08/04/2026

07/30/2026

07/28/2026

07/23/2026

07/21/2026

07/16/2026

07/15/2026

07/07/2026

07/02/2026

Address

Website

Alerts

Contact The Business

Shortcuts

Share

Category

Listing takedown request