this is honestly wild.
weโve been deep in the trenches building Poof V2 on the bleeding edge -
* AI models
* complex prompt orchestration
* crazy agentic flows
* massive context plumbing
* onchain infra
the goal was simple but brutal:
build the most capable vibe-coding engine for onchain.
and in the process... we built something much bigger ๐
Poof V2 is LIVE
Type a prompt. Get a full onchain app on Solana.
Not a frontend with a wallet button bolted on. Wallet auth, smart contracts, security scanning, backend, UI. All from your first prompt.
Think Claude Code, but purpose-built for Solana ๐งต๐
Hot take: the next serious agent product is not the one that feels most autonomous.
It is the one with the cleanest operating loop.
Start, observe, cap spend, verify, stop, resume.
The magic demo matters less than the boring controls.
Tencent's WorkBuddy Enterprise launch is a useful agent signal.
The hard part is shifting from super individuals to super teams.
Agents need shared context, reusable skills, controls, and a human quality gate.
Where does agent work become team infrastructure for you?
GitHub just made Copilot cloud agent callable from an API.
That is the real shift.
When a coding agent can be started by a script, portal, or release workflow, it stops being a chat tool and becomes build infrastructure.
What would you automate first?
Quick tip: ask the agent for the smallest patch that could prove the idea.
Not the final architecture.
Not the full refactor.
Not the beautiful generalized system.
One useful slice, one visible behavior, one way to undo it.
Agent memory is becoming infrastructure, not a feature toggle.
Walrus launched a portable, verifiable memory layer for agents across Claude, ChatGPT, Gemini, MCP, and SDKs.
The real question is ownership.
When an agent remembers your work, who controls the memory?
When you review AI-written code, which signal makes you slow down first?
1. diff touches auth
2. no test proves the risky path
3. summary sounds too confident
4. rollback is unclear
I trust speed more when the stop signs are obvious.
Solana just made subscriptions a native payment primitive.
That is boring in the best way.
Recurring API billing, payroll, agent spending limits, and stablecoin invoices should not require every team to rebuild billing rails.
What onchain product gets easier now?
My current agent prompt got better when I added a stop list.
Stop before:
touching auth
changing money flows
deleting data
expanding scope
guessing product intent
The point is not less autonomy.
It is putting the human at the expensive decisions.
The IDE was the easy place for coding agents to start.
The bigger shift is agents operating the loop around code: browser, terminal, screenshots, tests.
I wrote about why that changes both productivity and risk.
What would you let an agent touch outside the editor?
I wrote about the security shift I think builders are underrating:
AI made bugs cheap to find.
The hard part is now triage, patching, and judgment.
A scanner that outruns your response loop does not make you safer. It makes the backlog visible.
OpenAI and AWS just turned Codex into something enterprises can buy through the stack they already trust.
That matters more than one more model dropdown.
For agents, procurement, auth, billing, and governance are product features.
What makes an agent usable at work?
Local agents are getting their own PC shape.
NVIDIA and Microsoft announced RTX Spark Windows machines with OpenShell and new agent security primitives.
The interesting part is not the chip.
It is the control boundary.
Would you run an agent on your daily machine?
Hot take: the best AI products will not feel like magic.
They will feel interruptible.
What did the agent touch?
What did it skip?
What proof did it leave?
Where can I stop it?
Autonomy without those answers is just risk at speed.
OpenAI just pushed Codex further out of the IDE.
On Windows, Codex can now see, click, and type inside apps, while you steer it from mobile or a Mac.
The agent is starting to feel less like a coding sidebar and more like a remote operator.
What would you let it touch first?
I care less about Anthropic's Opus 4.8 benchmark bump than the workflow signal.
Claude Code can now plan a big task, fan out subagents, verify, then report back.
Coding agents are becoming review systems.
What would you trust an agent to parallelize first?
Quick tip: before asking an agent to add a feature, ask it what it would delete first.
If it cannot name dead code, redundant state, or a smaller path, the task probably needs more thinking before more implementation.
poofy is about to get dropped in some group chats tonight
Hotseat โ AI trivia, any topic, 10 questions, no install
what topic does your group deserve