WTF is Agent?
How an autonomous agent actually works, one component at a time, from my own build. The anatomy, the design decisions, what I deleted, what I measured, and a survey of what already exists.
- 1The anatomy of an agentAn agent is a model in a loop plus seven things around it. The loop, memory and toolbox are the spine; the rest is where I rebuilt the most.Jun 3, 2026
- 2Memory: remember, cite, forgetStore the raw text, keep facts small, derive everything else, and make every belief point back to its source so it can be traced and forgotten. The design is built; two journeys pass, but a memory benchmark is still missing.Jun 3, 2026
- 3The toolbox: one door to every toolOne door that holds every login, a typed catalog the model searches, and code instead of one tool call per turn. The agent never holds a secret.Jul 13, 2026
- 4The harness: a loop that survivesKeep the model's loop, change where it runs: durable, one active turn per conversation, interrupts as signals. Let the model steer and delete the routers, lanes and planners that tried to help.Aug 14, 2026
- 5Context and skills: what the model readsKeep everything, show the model a summarised view, store big results as files, and compact in two tiers. Put mechanical steps in skill scripts, because an agent's latency is its number of model steps.Aug 23, 2026
- 6Runtime and sandbox: where agent code runsOne execution interface with several light engines, inside a sandbox that has no network and no secrets. I built a microVM per session first and replaced it within a day, because a pure function should not need a VM.Nov 14, 2026
- 7Waking, asking, stoppingAn agent only needs a wake when its own reasoning is needed later; everything else is a notification. Asking is a durable pause, approvals are the same pause, and stopping is a clean result, not a crash.Aug 7, 2026
- 8Knowing it works: evalsA suite can cap every model by counting cases no model can pass. Exclude the unreachable, put an interval on the score, and make every scenario pass its own oracle. With that fixed, the journey suite separates the models: 72 of 81 for the 35B, 32 for the 8B.Oct 26, 2026
- 9What I deletedI deleted nearly as much as I kept, and the reasons repeat: the model did it better, it duplicated something, it sat in the wrong layer, or it was heavier than the job. The design is what survived.Jun 3, 2026
- 10AI is a label, not a thingAI is a stack of ideas: AI, machine learning, deep learning, language models. Ask which layer someone means.Nov 16, 2026
- 11Machine learning is fitting a curveMachine learning finds rules from examples by nudging parameters to shrink the error. Only data the model has never seen counts.Nov 16, 2026
- 12Neural networks are lots of tiny dialsA neural network is stacked weighted sums with a bend, trained by gradient descent. It is matrix multiplication, which is why GPUs matter.Nov 16, 2026
- 13Language models are autocomplete at scaleA language model predicts the next token. Its output is fluent and plausible but unverified, so give it sources and check what matters.Nov 16, 2026