SELF-HOSTED AGENT RUNTIME

Shihao Qi

AI Agent EngineerChongqing

I maintain Autarkia, a self-hosted agent runtime that runs on a single machine. Conversation is only the entry point — the real work happens in tool calling, long-term memory, and task scheduling. The question it has to answer is a concrete one: after months of unsupervised uptime, how does a machine still not fall apart?

github.com/hakituo ↗More about me →

subservices in one process: 31
clients on the same backend: 6
unit tests: 6,500+
change records over 182 days: 1,281

01Writing

Four Seconds on a First Visit2026-10-29The site always felt fast to me because my browser had a cache. A cold request from Beijing took 1.3 seconds to first byte and four seconds to finish, once twenty-four. On a small server in Hong Kong it now finishes in 0.3.#Networking#Deployment Why the Proactive Messages Read Like a Checkup2026-10-29The prompt said not to ask whether I had eaten, while the code filled a "meals" slot every day. I switched to a short greeting first with the real topic carried into the follow-up conversation, then taught the scheduler when to stay quiet.#Agent#Product#Scheduling The Synthetic Set Picked C. Forty-Four Real Questions Picked B.2026-10-27A synthetic evaluation set gave a clean answer to what granularity memory should be stored at. Real questions from real chats gave a different one — and along the way turned up three measurement problems.#Memory#Evaluation#Measurement#LLM Reply Timing: How a Median of Eight Seconds Was Wrong2026-10-27Calibrating how long a role takes to reply against real chat logs, the first pass pooled more than a year of data and got the conclusion backwards. With the method fixed, it turned out the system was missing an entire middle tier.#Agent#Product#Measurement

All posts →All tags →

02Notes

2026-10-27A Clean Ping Doesn’t Mean Big Packets Get ThroughZero packet loss and 11 ms latency, yet downloads crawled at 10–25 KB/s. Lowering the MTU to 1400 made them roughly 300 times faster.2026-10-27Answer the Ping on the Socket That Received ItThe Android client opened 159 WebSocket connections in a day, with 160 abnormal closes. Three small bugs stacked up, and every return to the foreground left one more zombie connection.2026-10-27Keep .bat Files ASCII-OnlyA few Chinese comments added to a launcher script, and it crashed on double-click with a string of chopped-up command names.2026-10-22Never Hand an Auto-Fixer a DirectoryI meant to clean one file I had touched; six test files I never opened were rewritten along with it.

All notes →

03Projects

AutarkiaUnder active development

Self-hosted agent runtime on one machine

Conversation is only the entry point — the capability lives in the tool layer, the memory layer, and the scheduler. 94 tools are isolated per role, long-term memory both forgets and distills, and six clients share one backend.

Python / FastAPI / C++ / llama.cpp / React / Kotlin / Swift / Docker

EdgeSchedResearch prototype · Thesis in progress

LLM residency, image bursts, and CPU backfill on a single GPU

A research prototype: on one GPU an LLM stays resident, image generation arrives in bursts, and CPU work backfills the gaps. The work covers admission control, migration cost, and crash recovery; the experimental results go into a thesis.

Python / C++ / CUDA / PyTorch / FastAPI / WebSocket / llama.cpp

All projects →

04Now

01Splitting the Android send pipeline and navigation, getting single files back under 400 lines.
02Layering tools per role: resident ones go into the schema every turn, on-demand ones are discovered by retrieval, and write operations clear a permission check first.
03Pushing proactive messaging further toward sending less: quiet hours, minimum interval, daily cap — the default is not to send.

05Site Stats

Posts: 17
Notes: 11
Projects: 2
Page views: —