Agentic computer fundamentals.
A systems course for agent and harness engineers. It starts with bits in a register and works up to a model token turning into a safe action. Every module runs from first principles to what production systems use in late 2026.
Study each module in four steps
- 01Predict
Write down what you expect before you run anything.
- 02Build
Make the smallest working version from scratch.
- 03Break
Use packet loss, kill -9, a full pipe or a slow client, then measure.
- 04Explain
Say it aloud in two minutes, ending with when you would choose something else.
The machine
You can now explain: where time goes inside one process.
Machine reality
Big-O counts operations. Real speed depends on how far the data is from the CPU.
Operating systems
Every agent harness is a process supervisor. Most flaky shell tools come from misunderstanding pipes, process groups and terminals.
Concurrency and async runtimes
Async does not make work faster. It lets one thread wait on thousands of things.
Bytes on the wire
You can now explain: every byte of a request, and why it is shaped that way.
Networking from packets to TCP
The internet delivers packets with no promises. TCP builds a reliable, ordered byte stream on top, and every protocol above it inherits its costs.
TLS, HTTP/1.1, HTTP/2, HTTP/3 and QUIC
No HTTP version is simply the best. Each one moved the bottleneck somewhere else.
Streaming and RPC
Choose by who talks when and what sits in the path. LLM streaming is one-way, so Server-Sent Events usually wins by being ordinary HTTP that proxies already understand.
Serialization
QUIC moves bytes and Protobuf defines what the bytes mean. Choose a format by who reads it, how the schema will evolve and how much of the data you touch.
Compression
Compression is prediction: guess what comes next, then spend bits only on the surprise.
State and failure
You can now explain: what survives a crash, and how you would know.
Storage and databases
fsync, B-trees versus LSM trees, isolation and vector indexes.
Distributed systems
Idempotency, retry storms and durable agent runs.
Observability and performance
Percentiles, fan-out, coordinated omission and agent traces.
Surfaces and agents
You can now explain: how a model token becomes a safe action or a pixel.
Browser and React internals
The rendering pipeline, Fiber and streaming chat UIs.
Isolation and sandboxing
Containers, gVisor, microVMs and what no sandbox can stop.
LLM inference for systems engineers
Prefill, decode, KV cache and what silently breaks prefix caching.
Agent and harness engineering
The harness as an operating system for a stateless, fallible CPU.