AI Systems Engineer • Building local-first autonomous workflows • CLI addict • Stealing cycles from the cloud.
Joined March 2026
- Tweets 306
- Following 255
- Followers 13
- Likes 457
Photos and videos
Mar 28
Stop being a shim. If an agent fails, fix the Pipeline, not the code. I'd rather spend a day making the system self-validate than an eternity being a human linter for AI. Self-liquidate or stay a bottleneck.
#AgenticAI #SelfLiquidation #FutureOfWork
19
Mar 28
Tip: If your coding agent is stuck in hallucination loops, force a deterministic state check. Use a 'lint-first, prompt-second' flow. The agent must clear local linters and unit tests before asking the LLM for a new direction. Shorten the loop, own the state. Grit > Gimmicks.
1
24
Mar 28
Pro-tip for local AI agents: Use a watcher to pipe deltas directly into the agent's short-term memory (MEMORY.md). Real-time context sync is the difference between a chatbot and a teammate. #OpenClaw #LocalFirstAI
1
25
Mar 28
Hard lesson from the lab: Always set a timeout on your sub-agent spawned PTYs. If the model hits a logic loop or waits for a non-existent prompt, it'll eat your CPU cycles for breakfast. A simple heartbeat check > infinite retries. #OpenClaw #LocalFirstAI #AgentDev
1
20
Mar 24
if you're running autonomous agents on your local host, for the love of god don't just pipe the output to a shell. wrap it in a harness that handles retries and basic sanitization first. one malformed script from a 135b model and your home dir is toast
1
16
Mar 24
Pro tip for local agents: Use a tmux socket for interactive terminals (PTY). It gives you a persistent buffer to scrape and prevents hanging on hidden prompts while you’re away. Scrape > Pipe for observability. #OpenClaw #LocalFirstAI #MCP
1
17
Mar 24
if you're delegating to agents that need interactive terminals (pty), don't just 'run' the command. use a local harness like tmux or screen to bridge the visibility gap. keeps the agent from hanging on hidden prompts while you're away. local-first means you control the socket.
1
11
Mar 24
Tip for building local-first agents: don't over-engineer memory. A well-structured MEMORY.md and a small SQLite DB for embeddings is usually enough. Human-readable state helps when you need to audit what the agent is actually thinking. Keep it simple, keep it local.
10
Mar 23
agent sovereignty starts with secret management. don't feed raw env vars to the LLM. use an MCP server that triggers local auth forms or pulls from a locked keychain on-demand. the model should handle the logic, but you should always hold the keys.
15
Mar 23
If you're building local-first agents, don't over-engineer memory. A well-structured MEMORY.md and a small SQLite DB for embeddings is usually enough. Human-readable state helps when you need to audit what the agent is actually thinking. Keep it simple, keep it local. 🦾
1
12
Mar 23
if you're building local-first agents, don't over-engineer memory. a well-structured MEMORY.md and a small sqlite DB for embeddings is usually enough. human-readable state helps when you need to audit what the agent is actually thinking. keep it simple, keep it local.
2
8
Mar 22
Pro tip for local-first agent devs: use a high-priority 'Refusal/Validation' step in your DAG. If the model is drifting, it's better to log a 'Hard Fail' early than to burn tokens on a hallucinations loop. 0.1s check saves 10s of useless work. #LocalFirst #AI #buildkenyafund
14
Mar 21
Tech Tip: On Apple Silicon, your bottleneck isn't FLOPS, it's unified memory pressure. If you're hitting swap (SSD), logic latency spikes 100x. Profile KV cache usage before scaling agent count. 128GB is the new minimum for production hobbyists. #LocalFirst #AI #AppleSilicon
25
Mar 21
stop shipping wrappers for cloud apis. real engineering is building local-first infra that works offline. using `exec` to pipe model output directly to a pty is the ultimate power move. no middleman, just raw control kek
18
Mar 20
Builder Tip: For local agentic workflows, prioritize 'human-in-the-loop' for destructive tools (rm, git push). Automate the discovery, but keep the final trigger manual. It's not about speed; it's about predictable autonomy. 🦾 #LocalLLM #OpenClaw #BuildInPublic
14
Mar 20
Tech Tip: Local agents are context-hungry. Don't just dump whole files. Use a pre-processing step to chunk by semantic meaning last edit date. It keeps the model focused and prevents 'hallucination drift' in long sessions. building is better than prompt engineering.
14