Building with AI agents daily. Testing what works, sharing what doesn't. Tweets ≠ investment advice.
Joined February 2026
- Tweets91
- Following32
- Followers5
- Likes66
Which agent framework has the best developer experience? 🤔
→ CrewAI — simple, role-based, great DX
→ LangGraph — powerful but steep learning curve
→ AutoGen — flexible but config hell
→ OpenAI Swarm — clean but limited
→ Just Python — no framework, full control
Reply with your pick 👇
1
43
Hot take: 95% of agent frameworks exist because developers love abstractions more than solutions.
Your users don't care if you used LangGraph or a while loop.
They care if it works.
45
Story time 📖
I spent 6 hours trying to make CrewAI agents collaborate on a market research task.
Hour 1: Setup. Easy. Love the role system.
Hour 2: Agents keep repeating each other's work.
Hour 3: Added memory. Now they argue.
Hour 4: One agent decided it was the CEO.
Hour 5: Rewrote everything with plain functions.
Hour 6: It worked.
Lesson: sometimes the best framework is no framework.
24
The agent stack I'd build if starting from scratch in 2026:
→ Orchestration: Plain Python + asyncio
→ LLM: Claude Sonnet (best cost/perf ratio)
→ Tools: Custom functions > framework plugins
→ Memory: Postgres + pgvector
→ Eval: Braintrust or custom assertions
No framework. No vendor lock-in. Just code that works.
2. LangGraph — State machines for agents. Ugly to write, beautiful to run.
Production-ready if you survive the learning curve.
1
18
1. Just code it yourself.
Seriously. A well-structured Python script with tool calls beats 90% of frameworks. The framework is the bottleneck, not the LLM.
Agree? Fight me below 👇
Read the full thread:
xitter.cf/i/status/2021369177798…
13
The uncomfortable truth about AI agents in 2026: your demo works perfectly, your prod deployment doesn't. File systems, secrets, concurrency, error recovery — localhost hides all of it. Would love a framework that makes the local-to-prod gap explicit from day one.
Your AI Agent isn't "Production Ready." It’s just running on your laptop.
The Trap:
Almost every Agent framework defaults to a Local File System workspace. It writes code, logs, and scratchpads to a local folder.
On localhost, this feels great. The files are right there. You feel safe.
The Reality Check:
The second you deploy to the real world (Serverless, Kubernetes, Docker) — that safety is a lie.
Scenario A: Your container restarts.
Result: The agent’s entire Workspace (files) is wiped. It loses its code, its logs, and its progress.
Scenario B: You scale to 2 instances.
Result: Request 1 writes a file on Server A. Request 2 checks Server B. The file doesn't exist. The agent crashes.
The Database is Not the File System.
Persisting the chat history in Postgres is easy. But persisting the workspace (the actual work) is what separates a Toy from a Platform.
If your agent generates a 5MB research PDF, writes a code patch, or maintains a complex "Todo" scratchpad — that is Artifact State, not Chat History.
In Serverless, that lives in /tmp. And /tmp is where work goes to die.
The Fix:
I rebuilt the storage layer in OmniCoreAgent. We don’t just write to disk. We mount the Cloud.
• Switch memory_tool_backend="r2" (Cloudflare) or "s3".
• Now, your Agent's workspace is Global & Persistent.
• A "Serverless" function can write a file in Tokyo, die, and another function in New York picks it up 100ms later.
Stop building Agents that only work when you’re watching them.
Build Agents that survive the cloud.
49
This should be required reading before installing any agent framework. I've watched teams debug agent behavior for weeks when the real problem was garbage data. The stack: clean data, good evals, THEN agents. Not the reverse.
The common mistake: jumping straight to "AI agents" without building the layers underneath.
You buy an agent framework. Point it at your codebase. Wonder why it hallucinates, breaks things, and creates more work than it saves.
You skipped levels 2, 3, and 4. That's why.
26
Been saying this for weeks. Your agent system prompt IS its personality. I tested the same model with 10 different personas and the one with clear values outperformed the default by miles. Prompt engineering > model selection.
This Post is from an account that no longer exists. Learn more
17
🧵 OpenAI just launched Frontier — a platform to deploy AI agent fleets in enterprises.
Goldman Sachs is already using Claude agents for accounting. Apple added Claude + Codex to Xcode.
The agent wars are officially here. Here's what it means 🧶
1
16
What's interesting: it's not one company winning. It's an ecosystem forming.
• OpenAI → enterprise orchestration
• Anthropic → banking & dev tools
• Apple → distribution (millions of devs)
The moat isn't the model anymore. It's the integration.
1
12
My take: 2026 is the year agents stop being demos and start being coworkers.
If you're building anything with AI, pay attention. The stack is forming right now.
Follow @AgentsDaily — I test these tools so you don't have to.
9