// Writing
sudo make me a blog.
I had a brief moment of weakness and stopped paying attention — and that was enough to get scammed. Even though I told the attacker his process looked like a attack, he convinced me it was required for security. Within minutes my account was emptied, and the bank refused reimbursement because I had approved the transactions. I’m still convinced this kind of attack could be prevented.
AI can generate correct code at high speed—but it cannot fix incorrect assumptions. In fact, it exposes human weaknesses even more than before: flawed requirements and design decisions become visible faster because they are implemented without friction. The real challenge is not coding, but defining the right problem. Used properly, AI can also help verify requirements and decisions early—but human errors remain the root cause.
Most private LLM projects stall at simple automations like news summaries or inbox cleanup. But the real goal is a context-aware system that understands your state, integrates fragmented personal data, and actively supports daily decisions. The challenge is no longer model capability—it’s data access, orchestration, and building a unified personal intelligence layer.
The goal feature restored autonomy, but the real takeaway was architectural: use Opus for high-cost reasoning and validation, Sonnet for execution. Attempts to scale with Fable failed early due to token limits, making it impractical for sustained workloads under current constraints.
I assigned Claude Code a large task with a fixed goal and forced it to stay on course. Even with nested Ralph Loops, a queue, and finally an hourly cron job, the agent repeatedly stalled before completing the work. It runs for now, but the experiment exposed a laziness for LLM-based agents.
Agentic coding makes implementation cheap, but for many teams the real bottleneck is now trust: code can be generated faster than it can be confidently reviewed.
If someone asks me…
Secret-Service is a Python MCP server that routes problems through a pipeline of seven specialised LLM-backed agents using a blackboard architecture stored entirely in SQLite. Instead of committing to one solution, it fans out into parallel strategy branches, gates each plan through a Judge before execution, scores the results, and returns the winner. Over time, the system learns: successful strategies are embedded and recalled for future problems; failures become anti-patterns.
DevAgentOps runs LLM agents inside CI/CD with hard gates: analyze role-based use cases/stories, write testable acceptance criteria and verify for contradictions/gaps; design by mapping features to modules, planning, modeling (formal checks where possible), refining to implementation-ready specs, and creating reverse prompts; implement via human-piloted agent coding plus agent review, MR, unit/integration/E2E/role tests, then manual merge - validate behavior and loop back with evidence
The faster we move, the more we need the discipline we abandoned to move fast. Agentic Safety-Critical Development embeds AI agents at every stage — verifying before, validating after, and throwing back errors upstream instantly. This is not slower Agile. This is Agile grown up.