Architecture

How this thing is actually built.

A multi-agent communications platform running on a single spot instance, written up as fourteen deep dives — one per layer, plus the concerns that cut across all of them.

The rest of these docs describe what the product does. This section describes how it is put together: the four-stage campaign pipeline, the crew layer that decides who works next, the memory the agents read from, the API that bounds the model bill, the data layer that is only allowed to add to the schema, and the deploy path that is also the recovery path.

Each post is a real account of one layer: the decision, the alternative it rejected, and the bill it is still paying. Every one carries a diagram drawn for that layer specifically, and none of them claims a capability that has not been measured.

Note —These pages are rendered from the markdown in docs/blog/series/ at build time, not hand-copied. The prose and the code it describes ship in the same commit, so the write-up cannot quietly drift from the system.

The opening

What the market already does, where it stops, and why the gap between the tools turned out to be the product.

01It Takes Them Days to Prepare a Campaign →The market is full of tools that write posts. I went looking for the one that runs the whole line, and could not find it.

The layers

One post per layer, from the campaign pipeline down to the box it runs on. Each one is a set of decisions and the constraint that forced them.

02Nothing Exists Yet →The campaign pipeline is four stages, and every decision worth explaining lives in the boundaries between them.03The Orchestrator's Tools Are the Other Agents →A typed blackboard, five delegation tools, and a dispatcher that refuses illegal moves. How the crew layer decides who works next.04Three Tiers of Memory, One Write Path →The text, the vector, and the name of the model that made the vector have to move together, or the crew ends up confidently retrieving nothing.05The API Layer Is Where the Model Bill Gets Bounded →Limits keyed by what they cost, two clocks on every stream, and authorization that refuses what it does not recognize 06The frontend layer: turn every silence into a failure →A thin client over a streaming backend, where an unhandled server event is a compile error and a failed load can never render as an empty list.07Expand Only →The data layer is allowed to add to the schema and to do nothing else, and every other decision in it follows from that one rule.08A Cost Column That Is Allowed To Be NULL →The observability layer: one capture point, three sinks, and a number that refuses to be confidently wrong.09The Deploy Path Is the Recovery Path →One spot instance at six to eight dollars a month, where shipping a release and losing the server run the same code.10Four Kinds of Test, Because There Are Four Kinds of Wrong →The testing layer: unit tests check code, contract tests check a boundary you do not own, evals check behavior, and a control checks the eval.

Cross-cutting

The four concerns with no box in the architecture diagram, which is exactly why they are easy to believe you are doing.

11Agent Engineering Is Not a Layer →Six concerns with no box in the architecture diagram, which is exactly why it is easy to believe you are doing them.12Security Is Four Boundaries, and the Bug Lives in the Seam →Authorization that fails closed, queries scoped to a tenant, egress checked on resolved IPs, and a deploy role that cannot overwrite what it deploys past 13The Eval Has to Be Able to Embarrass You →Two suites, one deterministic mock, thresholds calibrated from a measurement, and the positive control almost nobody builds.14Every Decision in This System Has a Price Tag →Nine choices, the alternative each one rejected, and the bill each one is still paying.

It Takes Them Days to Prepare a Campaign →

The market is full of tools that write posts. I went looking for the one that runs the whole line, and could not find it.