LLM Agents on SPERIXLABS

LLM-Redactor: What Leaves Your Prompt When You Talk to a Cloud LLM

Wed, 15 Apr 2026 00:00:00 +0000

Every time a coding agent sends a prompt to a cloud LLM, the full content of that prompt — your code, your credentials, your customer names, your internal project codenames — lands on someone else’s server. It may be logged, retained for training, produced in response to subpoena, or exfiltrated in a breach. TLS protects the wire. Nothing protects the content.

We built LLM-Redactor to measure exactly how much leaks and what you can do about it. The paper evaluates eight techniques on a common benchmark. This post is the practitioner’s summary.

Local-Splitter: Cutting Cloud LLM Costs by Putting a Small Model in Front

Wed, 15 Apr 2026 00:00:00 +0000

Cloud LLM tokens are expensive. Not in the “my AWS bill is high” sense — in the “I’m burning $0.015 per 1K output tokens and my coding agent sends 200+ requests per session” sense. Most of those requests don’t need a frontier model. “What does this function return?” doesn’t need Claude Opus. “Add a docstring here” doesn’t need GPT-5.

Local-Splitter is an open-source shim that sits between your coding agent and the cloud. A 3B parameter model running locally on Ollama triages every request: trivial ones get answered locally (zero cloud tokens), and complex ones get their prompts compressed before forwarding. The paper is now on arXiv: 2604.12301.

Resilient Write: Giving Coding Agents a Write Path That Doesn't Break

Sun, 12 Apr 2026 22:00:00 +0000

If you’ve spent any time watching an LLM coding agent work, you’ve seen it happen: the agent generates a perfectly good file, calls Write, and… nothing. The content vanishes. The agent retries the exact same payload. Five times. Then it gives up or cobbles together a cat >> file.tex <<EOF workaround in the shell.

This happened to me in April 2026 while an agent was producing a telemetry report. A LaTeX document containing redacted HTTP headers like Authorization: Bearer sk-ant-oat01-{REDACTED} got silently rejected by the host tool’s content filter. The prefix sk-ant- was enough to trigger the regex. No error. No feedback. Just silence and wasted tokens.