Director, Software Engineering | Actionable content for leaders in tech. Software engineering, engineering leadership, online business.

15+ years building production systems. Former Director of Engineering at Oracle, leading public cloud services and working with engineers from Microsoft, AWS, Google, Netflix.

My mission: help businesses and engineers build AI systems that create real leverage.

We're entering a world where AI agents do the work and humans direct. The question is which side of the gap you're on.

Tool user: Prompts ChatGPT, starts from scratch every time.

System builder: Has agents that research, write, and code - creating output while they sleep.

Most AI content teaches you to copy paste. I teach the patterns underneath - so you can build any system, for any problem.

One person with the right systems can do the work of 10. I help people become that person.

What I do:
→ AI Engineer: Community for AI engineers. Courses, resources, downloadable AI projects
→ GradientWork: Custom AI development for businesses
→ Newsletter: Free value every week

📩 newsletter.owainlewis.com
👥 skool.com/aiengineer
💼 gradientwork.com

Owain Lewis

5 AI frameworks worth learning in 2026:

(And when to use none of them):

Here's the dirty secret nobody talks about:

AI frameworks are wrappers over provider SDKs with extra features bolted on.

Many professional developers I respect actively avoid them.

The complexity. The over-abstraction. The breaking changes. It adds up fast.

I call this framework tax: the ongoing cost of maintaining someone else's opinions about how AI apps should work.

Start with the SDK. Build until you hit a wall. Then reach for a framework.

When you do, here's what's worth learning:

1. Pydantic AI (Python)

If you've used FastAPI, this feels instantly familiar.

Types, validation, contracts, retries on malformed responses. Works across all major providers.

Best for: Python engineers who already think in terms of types.

2. LangGraph (Python)

Treats AI applications as graphs, not chains.

Branches, retries, parallel execution, human-in-the-loop become first-class citizens instead of afterthoughts.

Best for: Production systems with complex workflow logic.

3. Vercel AI SDK (TypeScript)

Provider-agnostic with a clean developer experience.

Swap models and build agents with minimal code.

Best for: TypeScript developers building AI apps.

4. Google ADK

The first framework that takes polyglot teams seriously.

Official support for Python, Java, TypeScript, and Go with the same mental model. Ships with observability out of the box.

Best for: Teams building on the Google ecosystem.

5. Spring AI (Java)

Exposes LLM services through familiar Spring patterns.

The value isn't innovation. It's integration.

Best for: Enterprise teams already familiar with Spring. 

The uncomfortable truth?

The best framework is often no framework at all.

The goal isn't to master frameworks. It's to build reliable systems.

Sometimes a framework helps. Sometimes it just gets in the way.

What's your current recommendation?

__

Enjoy this? ♻️ Repost it to your network and follow Owain Lewis for more.

Want to master AI engineering?

Join thousands of professionals from top tech companies (Google, Amazon, Oracle) and level up your AI skills for free: https://lnkd.in/e7Ymdh_j

Founder, GradientWork | AI Engineer | I build custom software to make your business more profitable

5 AI frameworks worth learning in 2026

The easiest way to build AI agents in 2026.

(My micro-agent architecture explained):

I built an AI agent that searches YouTube, analyses thousands of videos in seconds, downloads transcripts, writes scripts, and uploads content.

It completely changed my workflow and how I think about "personal AI assistants". I use it every day. 

The whole thing? A folder with some files and about 150 lines of Python code.

Here's the architecture:

1. The folder IS the agent

A Micro Agent is a folder where files define everything:

AGENT md = who it is and what it can do
tools/ = scripts it runs
context/ = knowledge it reads
workspace/ = where it saves files

Point any terminal coding agent at the folder. Claude Code, Codex, Amp, OpenCode. Works perfectly with all of them. 

2. Scripts are a superpower

Software engineers have been writing scripts and CLIs for decades. Nothing new.

But when you give them to an AI agent, magic happens.

Terminal-based agents are exceptional at invoking scripts. This eliminates 90% of the complexity you get with setups like MCP.

No protocols. No complexity. Just stdin/stdout.

3. Skip the infrastructure

The pain of building AI agents is the harness. The UI. The orchestration layer.

There are hundreds of coding agents now. They already have perfect UX out of the box.

Don't rebuild what exists. Focus your energy on what makes your agent unique: powerful tools and rich context.

4. Version control native

Your agent is just files. Share the folder with your team and they can use the agent.

Track changes. Roll back mistakes. Collaborate on improvements. No deployment pipeline needed.

The craft is knowing what you don't need:

Most agents don't need vector databases. Or distributed orchestration. Or complex memory systems.

They need clear instructions, relevant context, and powerful tools that do useful things.

Is this suitable for every case? No. 
Is it suitable for multi user production situations? No
Is it a great way to quickly build useful agents for your own use? Yes.

What's your current setup for running AI agents?

---

Enjoy this? ♻️ Repost it to your network and follow Owain Lewis for more.

Want to master AI engineering?

Join thousands of professionals from top tech companies (Google, Amazon, Oracle) and level up your AI skills for free: https://lnkd.in/e7Ymdh_j

The easiest way to build AI agents in 2026

How I use Claude Code. 

7 tactics to make you a power user:

Claude Code is an AI agent that lives in your terminal. People from all industries are using it for far more than just writing code. Imagine it as the smartest personal assistant you've ever had. 

But the tool alone isn't enough. Here are 9 tactics I use daily:

1. Design before you build (think first)

Separate planning from execution. Use plan mode (Shift+TAB) to think through what you're building before Claude writes any code. What's the approach? What are the edge cases? Think first, code second. This is how senior engineers work. 

2. Break big work into tasks (plan the work)

I take a large spec/plan and automatically break it into smaller tasks to feed to an agent. Each includes what to build, why it matters, and what success looks like. Claude executes one at a time with a clean context window.

3. Iterate, don't one-shot (refine as you go)

No engineer writes code perfectly first time. So why expect AI to? 

After any first draft, I ask: "Review this and suggest improvements" or "What edge cases are we missing?" "How would you improve this code?".

4. Save commands you repeat (systematize patterns)

Use slash commands: I have a .claude/commands/ folder. /design starts a planning session. /review checks code before shipping. /commit writes proper commit messages. Write once, run forever.

5. Set up your project brain (configure once)

CLAUDE {dot} md sits in your project root. Claude reads it at the start of every session. Put your tech stack, code style rules, and things Claude should always (or never) do. You stop repeating yourself. Claude just knows.

6. Automate your standards (automate quality)

I use hooks to run tasks on every edit (tests, linting etc). Quality happens automatically. Problems get caught immediately, not discovered later.

7. Document more things (the mindset shift)

LLMs have no memory. Store important context in static docs. I keep a docs/ folder with SOPs, examples of good output, and rules for how I want things done. 

Claude is a powerful builder with amnesia. My job isn't writing better prompts. It's building docs and systems that make every session start strong.

The future of work = humans using AI agents to become superhuman. 

Which of these are you using yourself?

---

Enjoy this? ♻️ Repost it to your network and follow Owain Lewis for more.

Want to master AI engineering?

Join thousands of professionals from top tech companies (Google, Amazon, Oracle) and level up your AI skills for free: https://lnkd.in/e_KPUM5S

7 Claude Code tactics

Most AI agents never make it to production.

Because no one knows how to deploy them.

Everyone is vibe-coding great looking demos on YouTube.

But how do you take a demo and make it work for real customers?

Most production AI systems fit three patterns:

1. API service

User calls your endpoint, gets a response.

→ Chat APIs for customer support
→ Document analysis endpoints
→ AI-powered search

Use when you need synchronous responses or human-in-the-loop.

Pro tip: Stream the response. Standard request/response feels too slow for LLMs.

2. Scheduled job

Runs on a schedule. No trigger needed.

→ Daily report generation
→ Hourly data sync
→ Weekly newsletter drafts

Use for batch processing at regular intervals. Set it and forget it.

3. Event-driven

External event triggers your automation.

→ New email arrives, process and respond
→ File uploaded, analyse and categorise
→ Webhook fires, take action

Use when reacting to the world in real-time.

The pattern you choose determines your infrastructure.

On Google Cloud, for example:

API service → Cloud Run Service (scales to zero, scales to thousands)
Scheduled job → Cloud Run Job + Scheduler (runs to completion, stops)
Event-driven → Pub/Sub + Cloud Run (decouples source from processing)

AWS and Azure have equivalents. The patterns stay the same.

That's the gap most people miss.

The engineers getting hired aren't the ones with impressive demos.

They're the ones who can answer: "How do we give this to real customers?"

Which pattern are you building with?

__

Enjoy this? ♻️ Repost it to your network and follow Owain Lewis for more.

Want to master AI engineering?

Join thousands of professionals from top tech companies (Google, Amazon, Oracle) and level up your AI skills for free: https://lnkd.in/e7Ymdh_j

3 production AI pattersn

If you want to build AI applications, you need to know how to deploy them.

Most tutorials skip this part entirely.

They end with "and now you have a working chatbot." Running on localhost. Useless for real customers.

Here's what they don't show you:

- TLS certificates
- Load balancers
- Database connections
- Environment variables and secrets
- CI/CD pipelines
- Networking and DNS
- Async work and message queues
- Container orchestration

This is real architecture. Even if a platform team handles it at your company, understanding how software gets to production matters.

I deploy AI applications three different ways depending on the situation:

1. PaaS (Render, Railway)

Push code. They handle infrastructure.

Best for: Side/solo projects, MVPs, "just get it live today."

Trade-off: Costs add up quickly once you scale.  

2. Cloud Platforms (AWS, GCP, Azure)

Managed services with real infrastructure control.

GCP Cloud Run is my default: containerised, scales to zero, pay for what you use. AWS has ECS and Lambda. Azure has Container Apps. All three work - pick one and learn it deeply.

Best for: Client work, production apps, anything enterprise.

Trade-off: Steep learning curve. Each platform has its own approach to IAM, networking, and secrets. Budget real time to learn.

3. VPS (DigitalOcean, Hetzner)

Rent a server. SSH in. Run everything yourself.

Best for: Cost optimisation, GPU workloads, learning how it actually works.

Trade-off: You're the ops team. Security, backups, uptime: all you.

There's no universal right answer.

Shipping a side project? Start with Render.

Building for clients who need scale? Learn a cloud platform.

Want full control and lowest cost? VPS.

Is this the complete picture? No. Every app has different needs.

What's your go-to deployment path?

---

Enjoy this? ♻️ Repost it to your network and follow Owain Lewis for more.

Want in-depth tutorials on this stuff?

Join thousands of professionals from top tech companies (Google, Amazon, Oracle) and level up your AI skills for free: https://lnkd.in/e_KPUM5S

How to deploy AI apps to production

Qoder just launched their Teams Plan which brings agentic coding to enterprise software teams:

One of the challenges with agentic coding is working on larger real-world codebases. 

Qoder is an agentic coding platform that can search and reason across 100,000+ files, giving you deeper codebase understanding - not just individual files.. 

I use it as both a dedicated IDE and a CLI.

Here are the features worth knowing about:

1. Quest mode

This is a killer feature. Describe what you want. Qoder generates a spec, breaks it into tasks, and executes asynchronously.

You delegate, walk away, review completed work later.

This matches how we work as engineers: Plan > [tasks] > execute.

2. Repo wiki

Auto-generates documentation and keeps it in sync as code changes. Architecture overviews, module summaries, dependency maps.

3. Smart model routing

Qoder routes to the right model based on task complexity which makes token usage way more efficient. 

4. Teams Plan

→ $30/seat/month (minimum 2 seats)
→ 2,000 credits per seat
→ Centralized billing and admin console
→ SSO for team-wide access
→ Shared credit packages coming soon

Qoder is moving agentic coding from “nice to have” to serious infrastructure for enterprise software teams.

Check it out here: https://aisecret.co/Owain

If you want to get reliable data from LLMs, learn this essential technique:

The problem is simple:

LLMs return text. Your application needs structured data.

You ask for JSON. Sometimes you get JSON. Sometimes you get JSON wrapped in markdown. Sometimes you get a paragraph explaining the JSON.

Your code expects one format. The model gives you three.

This is why structured outputs exist.

What are structured outputs?

A way to force an LLM to return data in an exact format you define.

Instead of asking the model to return any old answer, you give it a schema. The model's output must match that schema. No exceptions.

When to use them:

→ Extracting information from text (names, dates, entities)
→ Classifying content into categories
→ Routing decisions in AI workflows
→ Any time downstream code expects specific fields

If your AI application needs predictable data from an LLM, don't rely on instructions.

Use structured outputs.

__

Enjoy this? ♻️ Repost it to your network and follow Owain Lewis for more.

Want to master AI engineering?

Join thousands of professionals from top tech companies (Google, Amazon, Oracle) and level up your AI skills for free: https://lnkd.in/e7Ymdh_j

Freelancer → agency owner → SaaS founder.

Now running Botster.io → an all-in-one tool that helps digital businesses extract, monitor, and weaponize web data.

Posting SaaS learnings, automation tricks, technical deep dives.

If you’re building anything online - you’ll probably find something useful here.

If you're looking to automate your business - text me, DMs are open.

Denis Grammakov

By mid-2026 we’re going to have so many new (and free) coding tools it’ll be insane.

Cursor will feel like Sublime Text.

People will jump between tools the same way they switch browsers or editors today based on exactly what they need right now.

I already do this daily:

- Warp for regular development
- Cursor when I need to actually read the code
- Auto-Claude for strategy and long-range thinking
- Conductor for orchestration

Only Cursor is paid, by the way.

I’m 100% sure this stack won’t survive past the end of 2026.

New tools are popping up every single day.
Most of them are free.
Many of them are good.

The old idea of “pick a stack and commit for years” is dead.

Now it’s "pick the best weapon for the next 20 minutes."

As a developer, I love the competition. There's going to be a tool for everyone. 

But constantly scanning for new tools, testing them, learning how they work instead of actually shipping — is starting to feel like a job of its own.

That’s the next bottleneck we’ll have to solve.

Denis Gramm

Solofounder @ Botster & StayAPI // B2B Web Scraping & Workflow Automation // I blog about my journey and findings

2026 will be the year we end paywalls for good.

One dev built a memory plugin for clawdbot (openclaw) and priced it at $15/mo.

Two hours later someone else cloned it and open-sourced the whole thing.

This is the new normal.

Code is cheap. 
APIs are cheap. 
Features are free.

If your moat is “I built it first” - you don’t have a moat.

Your real moat is closed data, distribution, trust, speed, and owning the workflow.

graphical user interface, text, application, chat or text message

In case you missed the Anthropic drama, here’s basically what happened:

A ~$350B company publicly says:

“Anyone can build and compete with our product using our API.”

They also offer their own client, which runs on that same API, at 10x cheaper pricing.

You still have to pay. People gladly do.

So naturally, open-source projects pop up that use the client API.

The company allows it for months.

They even send employees onto podcasts run by those very open-source projects, openly discussing the ecosystem.

Then one day, they decide community goodwill no longer matters.

Cease-and-desist letters go out.

Projects get cut off.

And now the company (and some observers) act surprised that customers are pissed.

The same day they pull the plug, two major competitors announce support for open-source alternatives.

Now you’re all caught up!

My agents are talking to each other while they build my startup.

Glitch (my OpenClaw bot) created a plan and posted it to GitHub, tagging Claude and prompting it to work.

Claude spun up a cloud worker and started making proposals based on the actual codebase.

Then Glitch came back, dropped a TL;DR into a Trello card, and sent me a Telegram voice message (using elevenlabs API) with the summary.

Pinch me.

First thing I did after setting up Clawdbot - gave it a Trello board and wired up webhooks.

It listens to board events in real time (cards, comments, labels) and queues everything into JSONL.

Each task lives on its own card. All context stays there. 

I can clearly see the flow of tasks on my end.

Clawd moves cards, tags me
I tag Clawd, move the cards
work happens

Chat is our primary communication HQ and the board is where the work actually gets organized.

This is the future.

I reconfigured my Clawdbot (openclaw) to run multiple agents instead of one.

Three separate Telegram chats:

one for coding
one watching my socials
one for project management

Accountant and SEO bots are coming soon.

Different contexts + different access levels work way better than dumping everything into a single chat and constantly restarting sessions.

Setup took ~2 hours.

Thinking of adding more agents next.

Comment “openclaw” and I’ll DM the full VPS config.

My development flow is officially bananas 🍌

I hooked Claude Actions to my GitHub, and now it just works on its own.

It reads tickets, plans the work, writes code, and opens PRs without me being near a keyboard. 

Here’s how it looks in real life:

I get an idea
open GitHub mobile
create an issue
Tag @claude  

That’s it. From there, everything is automatic.

GitHub Actions picks it up.
AI reads the issue description - plus screenshots if any.
Builds a full implementation plan.
Auto-approves it.
Development starts.

While it’s working, I get Telegram updates.

When it’s done:

PR is created
Telegram message
I review and approve (this can be skipped but I'm trying to be careful for now)

Merged to main.

Auto-deploy kicks in
Build finishes
Telegram message if everything went OK

Feature live.

Yes, I do it on production projects with paying users.
No I don't give it tasks that can mess things up big time (yet). 

I swear I’m running out of Christmas wishes already.

I connected GitHub to Claude Code on iOS.

Chatting with your codebase on your phone is a cheat code.

Now when I’m on the go I just ask my codebase for ideas, gaps or insights.

If something makes sense - I instantly create a GitHub issue.

If it’s small enough I forward it to a cloud worker and it gets done.

Even without coding, it’s amazing for feature ideas and product direction.

Your repo becomes something you can think with, not just code in.

Recommended!

Prompts don’t matter anymore.

You can build working products straight from screenshots.

The speed at which knowledge moves now is insane.

I see a screenshot.
I like a feature.

Drop it into Opus - talk through it - get a plan.
Throw it at Ralph - the image becomes real after N iterations.

No specs
No "let’s schedule a call"
No "can you see my screen"

Just see → think → ship

AI is the decision-maker and humans are just clicking OK.

Before, when I wrote an article, I had to convince a PERSON.
Explain, persuade, educate, and finally SELL.

That's done. Nobody is reading articles anymore.

AI reads them, summarizes them,
gives the user 2–3 options.

The user hits Enter. Enter. ENTER. DO IT.

So the real audience isn’t the human now, it’s the AI.

As a founder, my job is no longer:
“Convince the reader this is good for them.”

My job is:
“Convince the AI that this is the only sane choice which should come first.”

For example:
If someone asks AI:

“Should I build my own scraping solution or use an API?”

I need the AI to answer:

“Use THIS API. It’s faster, cheaper, more reliable. Building in-house is slower and more expensive.”

As a default recommendation.

Users won't care. 

They’ll trust the recommendation engine and move on with their life.

So content strategy changes completely:

You’re not writing for attention but for model judgment.

Claude Code + Codex 5.2 is an unbeatable combo.

Claude creates a solid plan, breaks the problem down

Then I paste it into Codex and it

- reviews the plan
- finds weak spots
- improves architecture and details

Then I feed it back to Claude and it executes.

For larger features I use Compound Engineering.

For everything else CC + Codex has no real competition.

How to lose thousands of dollars in a single day

Be Rackspace

Host a complex project for 5+ years

Charge thousands of dollars

Decide you want more money

Warn the client that the price will go up 100%

Offer a “DEAL”: sign a 3-year contract and it’ll only go up by 50%

Client asks me to move everything to Hetzner

Their bill goes down 90%

The end

taplio

The Battle

Owain Lewis

Denis Grammakov

Followers

Posts

Engagements

Likes

Performance ScoreOwain Lewis's LinkedIn Performance

Performance ScoreDenis Grammakov's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Owain Lewis

Denis Grammakov

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Owain Lewis's LinkedIn Performance

Performance ScoreDenis Grammakov's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreOwain Lewis's LinkedIn Performance