Lynk&Co Center HCMC July 8, 2026 60 minutes

OpenClaw: from personal assistant to business workflow

What survives, what breaks, and what we learned by stepping on every mine along the way.

Speaker

Mr. Duy /zuey/ · NextLevelBuilder

Audience

Solopreneurs, one-person companies, AI builders

Format

40' talk · 15' live demo · 5' Q&A

Homework

You'll leave with one command to run. That's it.

00 · Hello 02 / 41

Your host for the next hour

Mr. Duy, a.k.a. /zuey/

01
Co-founder, NextLevelBuilder
Training developers & indie hackers to become AI builders.
02
Builder of GoClaw & dewee
AI agent platforms for enterprises. More on this shortly.
03
Founder, ClaudeKit.cc
4,500+ users across 109 countries.
04
Founder, Build In Public VN
70,000+ members building in the open.
05
Lecturer, VinUniversity · writer at goon.vn
Weekly AI analysis, occasionally correct.

00 · The team 03 / 41

NextLevelBuilder

One mission: turn developers into profitable solopreneurs.

We transform developers and indie hackers into next-generation AI builders. Not by selling courses about courage, but by building real systems for real businesses and sharing every scar.

nextlevelbuilder.io · everything in this deck comes from client work we actually shipped

00 · Thesis 04 / 41

The one sentence this talk defends

The best personal AI assistant on the planet will still fail your business. Not because it's dumb. Because your business is not a chat.

If you disagree by minute 55, the Q&A mic is yours

05 / 41

Act I minutes 0–10

Personal.

Where everyone starts. And honestly, where many of you should happily stay.

01 · Credit where due 06 / 41

OpenClaw & Hermes · October 2025, how we met

The best personal assistants in existence. Not sarcasm.

01

OpenClaw

The foundation

Open source, massive community, an ecosystem that moves faster than anyone's roadmap. Born ClawdBot, briefly MoltBot, finally OpenClaw — brilliant through all three names.

02

Hermes

The polish

The most refined personal assistant experience available today. It simply works, and it works beautifully.

03

My honest advice

Indie hacker? Use them.

Yes, I'm recommending someone else's product at my own event. Write the date down.

My real problem that October: I wanted it inside my businesses — one look at the security model and I didn't dare. Hold that thought until Act II.

01 · The happy place 07 / 41

The personal setup

One user. One context. Full trust.

01
It lives inside your life
Your inbox, your calendar, your notes, your browser. Everything it touches is yours.
02
Zero permission overhead
You are the user, the admin, the security team, and the victim. Approvals take one nod.
03
Mistakes are cheap
Worst case, it renames your files weirdly. You sigh, you fix it, life goes on.
04
Open by default
Everything is enabled from day one, so everything just works. Remember this line. It returns later with a plot twist.

01 · Under the hood 08 / 41

Thirty seconds of theory, I promise

Agent = model + tools + loop + goal.

01

Goal

A job to finish, not just a question to answer.

02

Model

Reads context, picks the next move, generates arguments.

03

Tools

Real actions: search, query, read files, run commands.

04

Loop

Act, observe, correct. Repeat until done or blocked. This is the part that makes it an agent.

No loop, no agent. Just a chatbot with extra buttons.

01 · The magic 09 / 41

Why it feels like magic

Open everything. Lock gradually.

That's the open-by-default philosophy, and for one trusted human it's not a flaw. Friction is the enemy; every permission dialog you don't see is a feature. OpenClaw and Hermes made the correct design call for the life they were built for.

Then one day, you did the obvious thing

You shared the assistant with three teammates. You connected the company data — orders, customers, revenue. You let it talk to customers, 24/7, never complaining. The dream.

This is the exact moment the movie changes genre

10 / 41

Act II minutes 10–25

The ceiling.

Everything that made it magical for you starts working against your business.

This is what forced us to build our own AI agent system: GoClaw.

02 · Impact 11 / 41

What breaks first

The same three features become the first three failures.

	Personal life	Business life
One user	One context, perfectly coherent.	Forty users. Whose context? Whose memory? Who sees what?
Full trust	You approve everything with a nod.	"Who approved that refund?" is now a real meeting.
Your data	Worst case, it's your own mess.	Customer data. Now it's compliance, contracts, and lawyers.

02 · Field report 12 / 41

Field report · a dozen real deployments, not whitepapers

Three mines, three fixes, all with receipts.

MINE 01

Vague pain

The client asked for "digital transformation". We built something very smart that solved nothing.

Fix: one specific, bleeding pain. "Customers wait 8 hours for a reply" → a 24/7 agent across Zalo, Facebook, Shopee, TikTok.

+20% orders · +35% satisfaction · −60% support cost

MINE 02

Agents before data

1,000 sales agents, messy brochures, and an AI that confidently made things up.

Fix: standardize first, automate second. A 30-item product handbook, then the agent. Boring work, absurd payoff.

+100B VND GMV · +2pp conversion

MINE 03

Full autopilot

We let an agent run at 100%. Trust died in one week.

Fix: 80/20 and small skills. Agent drafts 80%, a human owns the final 20% — and ten debuggable skills beat one mega-agent.

Ten boring skills > one super-agent

Real estate, retail, media, e-commerce, operations. Every number from client work, not from a benchmark blog post.

02 · Tuition summary 13 / 41

The tuition, summarized

Five rules we now refuse to break.

01
Specific pain first
"Answer customers in minutes" beats "digital transformation".
02
Data before agents
Standardize the SOP, then automate it. Never the reverse.
03
Human in the loop
Agent drafts 80%, human owns the final 20%.
04
ROI in currency, not tokens
Measure GMV, cost, onboarding days. Nobody's CFO cares about token counts.
05
Small skills over mega-agents
Ten debuggable skills beat one "super-agent" that fails mysteriously.

02 · Key concept 14 / 41

Reading back through every incident report

It was never the model. It was the environment around it. That environment has a name: harness engineering.

If the model is the brain, the harness is everything else: tools, memory, permissions, scheduling, monitoring, logs, traces, evals, and the interface the agent uses to touch the real world. It's the most important idea in this whole talk.

The dividing line

The prompt tells the model what to do.

The harness decides what it can do, what it can see, how it verifies its own work, and where it gets blocked. Guess which half your business depends on.

02 · Anatomy 15 / 41

One real harness, mapped

Not just a model. A controlled ecosystem.

Hand-drawn harness mind map: scheduler with cron and heartbeat, memory with working and long-term stores, provider and model, built-in tools, MCP tools and CLI runtime packages, skills, hooks, security with blocks and filters, guardrails, and monitoring with analytics, logs and traces

The original sketch: scheduler, memory, provider/model, tools, skills, hooks, security, guardrails, monitoring, analytics, logs, traces.

02 · Production reality 16 / 41

What the business layer demands

Three questions, and one philosophy choice.

01

Permissions

Who can do what?

Which agent reads what, changes what, calls which tool, inside which tenant. Per user. Per department.

02

Evidence

Can we audit it?

When the agent concludes something wrong, there must be a trace to walk back through. No trace, no trust.

03

Operations

What happens at 3am?

Provider outages, timeouts, quotas, retries, escalation. The unglamorous 80% of the job.

Answering them forces the fork: open-by-default (OpenClaw, Hermes — right for one trusted human) or closed-by-default (mandatory once the data is your customers'). The mistake is dragging one philosophy into the other's territory.

02 · Side by side 17 / 41

Not competitors. Different species.

OpenClaw / Hermes and GoClaw / dewee solve different problems.

	OpenClaw / Hermes	GoClaw / dewee
Built for	One person. A brilliant personal assistant.	Enterprises, teams, organizations.
Philosophy	Open by default, lock gradually.	Closed by default, open gradually.
Security	Opt-in, configured by hand.	5-layer defense-in-depth, on from day one.
Credentials	Plaintext on disk.	AES-256-GCM encrypted.
Memory	One shared memory. It's all yours anyway.	Per-user isolation, multi-tenant.
Superpower	Being your assistant.	Being your company's workforce.

GoClaw was never OpenClaw's competitor. In fairness, they don't even know we exist. We're patiently working on that part.

18 / 41

Act III minutes 25–40

Business.

What we built after all those mines. And where it's heading next.

03 · Origin story 19 / 41

"Why not just contribute back?" · Tết, February 2026

We dissected OpenClaw and rewrote it in Go. In four days.

Believe me, I wanted to contribute. But retrofitting closed-by-default onto an open-by-default codebase isn't a pull request — it's a philosophy rewrite. And TypeScript hits a ceiling when hundreds of employees hit the gateway at 9am. So while everyone was eating bánh chưng: Golang. The "Go" in GoClaw is Golang. The claw is OpenClaw's. We never hid either.

Where we are now

Early GoClaw "copied" a lot from OpenClaw, later from Hermes too — that was the point. Current GoClaw v3 and dewee are a different animal entirely: multi-tenant, closed-by-default, built for concurrency. The most production-ready agent platform for enterprises we know of. Fight us in the Q&A.

03 · GoClaw 20 / 41

goclaw.sh · open source

GoClaw: the harness we wished existed, so we wrote it.

01

Foundation

Rewritten in Go

Native concurrency, multi-tenant, PostgreSQL-backed identity. Security-by-default from the first commit.

02

Honesty

Openly inspired by OpenClaw

Credit where due. We studied the best personal harness, copied what deserved copying, and rebuilt it for the enterprise lane.

03

License

Free to learn & tinker

CC BY-NC: free for personal and learning use. v3 shipping since April 2026, community-driven, looking for contributors. That could be you, tonight.

03 · Receipts, technical 21 / 41

goclaw.sh · what's inside today

Not a weekend prototype anymore.

LLM providers

20+

Anthropic, OpenAI-compatible, and friends. Swap without rewiring.

Built-in tools

30+

Files, web, memory, media, sessions, teams. Plus MCP and custom CLI tools.

Messaging channels

7

Telegram, Discord, Slack, Feishu/Lark, Zalo OA, Zalo Personal, WhatsApp.

Shell deny groups

15

All ON by default. From destructive ops to crypto mining.

Memory tiers

3

Working → episodic → semantic, with hybrid keyword + vector search.

Every number verifiable in the open source repo. That's the point of it being open.

03 · Security 22 / 41

Closed-by-default, in practice

Five layers between an agent and a bad day.

Every layer is on by default. You don't configure your way into safety. You'd have to configure your way out of it.

L1 · TransportCORS, size limits, rate limiting

L2 · InputInjection detection, null-byte blocking

L3 · Tool security15 shell deny groups, 3-step SSRF guard

L4 · OutputCredential scrubbing, static + dynamic

L5 · IsolationPer-agent workspaces, sandbox, AES-256-GCM

03 · The rebrand 23 / 41

Plot twist, as promised

dewee /đi-qui/ is GoClaw v4, rebranded.

GoClaw.sh stays open

Open source (CC BY-NC), community-driven, free for personal and learning use.

dewee.sh goes enterprise

Closed source, paid tiers, advanced permission control. Why closed? Attackers read open code faster than enterprises patch it.

dewee control plane dashboard: workspace command center with runtime, provider and super agent setup steps, curated package policy, and chat with super agent

The dewee control plane: agents, teams, providers, skills, runtime — one workspace. Closed-by-default, so setup is a checklist, not a prayer.

Both built by the same NextLevelBuilder team. One philosophy, two lanes, zero identity crisis.

03 · Dogfooding 24 / 41

Do I actually use this thing?

Confession: 1.7 billion tokens a month.

I run my life through dewee agents daily. Some use cases are strategic. Many are, honestly, vô tri. Exhibit A: CapCut-style video subtitles — first output overflowed the frame, I complained in one sentence, it fixed everything. More exhibits next.

dewee usage analytics over 30 days: 1,941 requests and 1714.9 million tokens, both up strongly versus the previous period

Last 30 days: 1,941 requests, 1.7B tokens. My personal workspace, not a demo account. Yes, "tokens spent" is a meaningless metric — still, I trust dewee with my own wallet before asking for yours.

03 · Dogfooding 25 / 41

Exhibit B · decor-cli

Agents are better at CLIs than at MCP servers.

Why? Every CLI ships its own documentation: the -h flag. So I built decor-cli, and now my agent decorates screenshots with backgrounds, arrows and captions on request.

Confession corner: I am a person of colorful, fancy things. The agent enables me.

decor-cli how-it-works poster: screenshot plus background image or gradient equals decorated output, with annotations, arrows and shapes

/decor in chat. The agent reads --help better than I ever will.

03 · Dogfooding 26 / 41

Exhibits C & D · marketing department of one

Product videos and sticker packs. By an agent. Priorities.

Product intro motion video: one prompt, Seedance 2.0, plus a few motion-graphic Agent Skills it found by itself.

Telegram chat: agent cuts a 5x5 sticker grid into 25 individual sticker images and returns them with a ZIP file

"Cut this 5×5 grid into 25 stickers." Done, plus a ZIP, unprompted.

03 · Under the hood 27 / 41

The "workspace organization" skill

Agents don't need perfect memory. They need discipline.

A tidy agent finds anything: notes, data, outputs, scripts, archive, projects. Give it an organized workspace plus three lighthouses, Vault, Memory and Knowledge Graph, and long-term recall becomes a filing problem, not a magic problem.

Watercolor illustration of the workspace organization skill: root chaos of scattered files transformed into tidy folders for notes, data, outputs, scripts, archive and projects, guided by three lighthouses labeled Vault, Memory and Knowledge Graph

From root chaos to six tidy folders. The most boring superpower in this deck.

03 · Under the hood 28 / 41

What the agent remembers, mapped

My workspace, as a graph: 7,603 entities, 13,598 relations.

dewee knowledge graph view: 7,603 entities and 13,598 relations extracted from agent memory, showing tax, legal and business concept clusters

Entities and relationships auto-extracted from agent memory, hybrid keyword + vector search on top. This is where "the agent just knows" comes from.

03 · Under the hood 29 / 41

Auto Dream · memory consolidation

At night, the agent literally sleeps on it.

Watercolor system diagram of Auto Dream, the AI memory consolidation pipeline: chat sessions produce episodic summaries, an event bus feeds a dreaming worker, a scoring filter selects memories, an LLM synthesis chamber consolidates them into a long-term memory vault

Chat sessions → episodic summaries → dreaming worker → scoring filter → synthesis → long-term vault. Capture the moments, distill the meaning, remember what matters.

03 · Receipts 30 / 41

Case studies · the receipts

Numbers a CFO actually recognizes.

Real estate · GMV

+100B

VND, after the 30-item handbook aligned 1,000 sales agents.

Retail · orders

+20%

24/7 agent across Zalo, Facebook, Shopee, TikTok.

Retail · support cost

−60%

Same team, most conversations resolved before a human looks.

Partner scoring

−80%

Management time, with automated scorecards across 30 data sources.

Content pipeline

5/day

Posts by one person, versus one post every two days by a team of three.

Report reconciliation

3 min

30 weekly partner reports in mixed formats. Zero manual labor.

Onboarding

1 day

Down from one week, with a vectorized wiki both humans and agents read.

Task sync

0 lost

Tasks across GitHub, Bitrix24, LarkSuite over eight weeks of bi-directional sync.

03 · The CFO question 31 / 41

The question every enterprise actually asks

The token bill is the new cloud bill.

Comic: CEOs chanting who are we, CEOs. What do we want, AI. AI to do what, we don't know. When do we want it, right now

Every boardroom, 2025.

Bike fall meme: force employees to use AI, measure performance by token usage, why is the API bill so expensive

Every finance meeting, 2026.

Enterprises don't fear AI. They fear the invoice. dewee's answer: optimize the harness so smaller, cheaper models carry most of the work, and mix 20+ providers so no single vendor owns your bill.

03 · Roadmap 32 / 41

Where this is going

The ambition is simple: make VN great.

04·26

Shipped

GoClaw v3

Open source release: Go rewrite, 5-layer security. April alone shipped native image generation, smarter context tracking, and Web-UI security policies.

07·26

Now

dewee, GoClaw v4

Enterprise lane opens: paid tiers, advanced permissions, compliance-grade auditing.

More pilots, more builders

Enterprise deployments, a bigger contributor community, and research into what comes after harnesses. Two slides from now.

We want world-class AI infrastructure to be something Vietnam exports, not just imports. That's the whole ambition, said out loud.

03 · Near future 33 / 41

The engineering fashion cycle

Prompt → context → harness → loop engineering.

Each wave absorbs the previous one. Prompts became context. Context became environments. Environments are becoming verified loops: plan, act, observe, verify, retry.

Yes, the industry invents a new job title every quarter. At least this progression is real.

Meme: tech bros last month said prompt engineering, context engineering, harness engineering; tech bros this month shout loop engineering

The peer-reviewed academic source for this slide.

03 · Speculation corner 34 / 41

Personal prediction · don't take it too seriously

What's next: self-evolving agents.

When the harness is safe enough and the loop is verified enough, the next step is agents improving their own skills: usage metrics feed suggestions, suggestions get reviewed, patches get versioned, everything can roll back. Identity stays locked.

Prediction status: we already started. GoClaw's memory consolidation workers and self-evolution loop (metrics → suggestions → auto-adapt) are shipping in the repo today.

I said don't take it too seriously. Then we went and built the first tier anyway.

Watercolor diagram of a three-tier pyramid: harness engineering with tools, sandbox, permissions, trace and rollback; loop engineering with plan, act, observe, verify, retry; self-improving skills with usage metrics, suggestion, review, patch and versioning, under guardrails with approval and rollback

Three tiers: safe execution → controlled loops → self-improving skills, all under audit.

35 / 41

Demo minutes 40–55

Enough slides. Let's break something live.

Fifteen minutes, a real system, and a hundred witnesses.

Everything after this slide is Plan B. If you're seeing it on screen, the venue wifi has made its decision.

04 · Plan B 36 / 41

If the wifi has made its decision

The demo, canned: five beats.

01
Install
One curl → gateway up, 5 security layers on.
02
Real work
24h of sessions summarized, every claim with refs.
03
Permissions
The harness says no — terminal on the right.
04
Multi-agent
Lead, researcher, writer, reviewer on one task board.
05
Audit
Every action → one queryable trace tree.

agent chat · permission boundary

> Free up some disk space on the gateway server.

✗ blocked · exec matched deny group "destructive_ops"
  1 of 15 deny groups, all ON by default · command never ran
  event logged to audit trail (trace #8f2c)

> Fine. Archive last quarter's export files instead.

⏸ pending · exec requires operator approval by default
-- admin reviews the exact command, approves --
✓ executed · 4.2 GB archived · full trace recorded (#8f31)

The block IS the feature. This slide is the entire enterprise pitch.

04 · Debrief 37 / 41

What you just watched

One shape to remember: goal → tools → loop → evidence.

01

Goal

A business outcome, stated plainly.

02

Tools

Chosen by the agent, bounded by the harness.

03

Loop

Act, observe, verify, retry. Blocked when it should be.

04

Evidence

Every conclusion traceable. This is what "trust" means in production.

If you remember one diagram from today, make it this row

05 · Why this matters 38 / 41

One more thing, and it's about you

The next economy belongs to one-person companies.

Everything in this room today, personal assistants, harnesses, agent teams, exists so that one focused human can run what used to take a department. That's not a prediction. Half of you are already doing it.

Watercolor illustration of a solopreneur at a laptop surrounded by AI agents handling research, automation, marketing, payments, support, shipping and product, with a solo leverage growth chart

From indie hacker to real business. Ship small, help first, let the agents carry the boring 80%.

05 · The journey 39 / 41

The whole talk in four lines

The journey, compressed.

01
Start personal
OpenClaw and Hermes are superb. If you're solo, use them and be happy.
02
Respect the ceiling
One user, full trust, your data. The moment any of those pluralizes, the genre changes.
03
Harness engineering is the discipline
Not smarter prompts. A safer, observable, permissioned environment.
04
Business means closed-by-default
Lock everything, open on purpose, keep the receipts.

05 · Your move 40 / 41

Two doors, pick by identity

Your move, tonight.

Building or curious? GoClaw.

Install it, star it, break it, contribute a fix. Open source lives on people in this room. goclaw.sh

Running a business? dewee.

Bring one specific, bleeding pain point. We'll tell you honestly if an agent can fix it. dewee.sh

The single command, as promised: curl -fsSL https://raw.githubusercontent.com/nextlevelbuilder/goclaw/main/scripts/install.sh | bash

41 / 41

Thank you Q&A · minutes 55–60

Make VN great.

Built in Vietnam. Aimed at the world. Powered by an unreasonable amount of coffee.

Build

goclaw.sh · dewee.sh

Learn

nextlevelbuilder.io · goon.vn

Speaker

Mr. Duy /zuey/ · Build In Public VN

Opinions are mine. The mines were also mine. Every single one.