Identity management for agents is the thing nobody talks about but everyone hits. You don't notice it until you're 4 agents deep and realize you've spent your afternoon in IAM consoles instead of shipping. First-class concern. Full stop.
Full writeup with the Docker setup, browser fingerprinting, threat model, and what broke when the agent published this post through the pipeline.
rida.me/blog/why-i-f...
Separate email, dedicated 1Password vault, JIT secrets. Once it had a real identity, CAPTCHAs disappeared (Google OAuth), GitHub access worked (scoped account), and every workflow I built had auditable boundaries.
The hard problems of agent automation are identity problems.
The knock on OpenClaw is security. Fair. So I spent two weeks making it actually safe. Docker-first, Tailscale sidecar, Chrome in its own container. But the biggest unlock wasn't infrastructure. It was giving the agent its own identity.
OpenAI’s ChatKit ships a Python SDK. I’m using Rails.
I just built a bridge from MCP UI → ChatKit widgets.
When a tool returns a ui:// resource, Rails extracts the widget payload and streams it as an SSE event.
AI agents wrote most of the code after I defined the architecture.
Full Writeup:
What if technology didn’t feel so… hollow?
Some friends and I just released a manifesto about a world where tech leaves us feeling nourished (along with an evolving list of theses about how we can build it)
resonantcomputing.org
Context windows matter more than people realize.
I split agent work into focused sessions:
1. write the feature
2. run browser tests
3. fix the bugs with test output
It’s not elegant, but it’s efficient — and that’s the reality of coding with agents today.
rida.me/blog/setting...
Cursor made me think in file references.
Claude Code made me think in patterns.
Codex made me think in constraints and goals.
The evolution of agents is really the evolution of what you no longer need to explain.
Agents repeat a lot of unnecessary work.
I’ve started running “meta sessions” where the agent scripts any repetitive CLI chain into one command. It keeps context clean and makes the whole workflow more reliable.
Jobs > steps.
AI coding sessions succeed or fail long before the model starts generating code.
The real variable is the context you give the agent.
Modern agents don’t need step-by-step instructions — they need intent, constraints, architecture, and a definition of “done.”
Wrote up the full framework with failure modes for each pattern:
rida.me/blog/what-ma...
The thing I got wrong building these:
I thought the agent should prompt the handoff. "Ready for you to take over!"
It feels like the agent giving up. The user should claim control when ready.
That's what makes it a collaborator, not a tool.
The question that separates AI-native from AI bolted on:
Not "does it have AI features?" but "how does control flow between user and agent?"
The best apps transfer control progressively, conversation to drafting to polish, and the interface matches where you are.
MCP's auth model has a chicken-and-egg problem.
You can't discover tools without auth. You can't auth contextually without knowing which tools matter. You can't know what matters until the user asks.
My workaround in the image. But the spec needs a real answer.
So I built isolated dev environments. Each feature branch gets its own containers, database, tunnels, secrets.
When the blast radius is zero, you can finally let go.
Wrote it all up here: rida.me/blog/yolo-mo...
The problem wasn't the model. It was their environment.
Shared databases. Conflicting ports. State bleeding across branches. Of course I couldn't trust YOLO mode.
I spent 12 months coding with AI agents daily. Cursor, Claude Code, Codex.
The productivity was real. So were the failures. I kept seeing the same pattern: agents declare victory before the job is done.
Happy to share that I’ll be speaking at @confooca.bsky.social 2026 in Montreal!
Two talks this year:
• Agentic Coding: Building Features with AI Teammates
• Safe Agentic Dev Environments
If you're into AI workflows, coding agents, or dev tooling, would love to meet folks there.
Built a little tool called BranchBox. Every feature gets its own fully loaded and isolated dev environment. Worktrees, devcontainers, Docker networks, databases, ports, env vars. No clashes.
Great for humans. Even better for coding agents.
Repo: github.com/branchbox/br...
Current mode: 3 projects in parallel
5 Codex CLI tasks
2 Codex web tasks
3 Claude Code CLI tasks
1 Claude Code web task
Feels like speed chess! Fast moves, limited time, full focus.
Good thing I’ve got a larger context window than those agents 😏
Best part?
./bin/feature-teardown oauth
Cleanly removes:
- Worktree
- Container
- Database
- Tunnel
Ship aggressively. Clean up instantly.
YOLO mode with an undo button. /2
Would this change your workflow?
When you're juggling multiple Claude coding sessions and your local env becomes the bottleneck:
I built something for a safer YOLO mode:
./bin/feature-start oauth
→ Isolated worktree + container + DB + live URL
→ Ready for Claude/Codex
→ 10 seconds
Zero conflicts. /1
I missed Heroku’s magical Review Apps after moving to my own Hetzner box, so I rebuilt them with Kamal + GitHub Actions + Postgres schemas.
“/deploy” on a PR 👉 spins up an isolated env & subdomain
Closing the PR 👉 tears it all down
Full walkthrough ↓
rida.me/blog/kamal-g...
OpenAI released Operator and Computer Use Agent today.
I really like the "take control" feature and human-in-the-loop.
The fact that it relies solely on the screenshot with no page markup is impressive too. I needed both when building Pair Browsing.
Excited to try CUA when it comes out!
Long time indeed! Glad to be here. Hoping to reconnect with the community and it's so heartwarming to start with you <3
Inspired by browser-use (ported its code for a browser extension), DoBrowser fr the core idea, and Google’s Project Mariner.
If you’re curious, here’s the repo: github.com/rbarazi/pair....
More insights soon!
I’ve been tinkering with something I’m calling Pair Browsing. Think of it like pair programming, but for the web—a little AI agent that helps you navigate day-to-day browsing. Check out the quick demo below! 🤖
