A field guide to agentic engineering
Ten concepts and four longer pieces, worked end to end against a real record.
The subject here is harness thinking: the working idea that the engineering lives in the rings around the model — the loop that runs it, the context it is handed, the gates that check it, the record it leaves — and not inside the model itself. The ten concepts are stated so they are useful anywhere, whether your harness runs for a minute at your elbow or all night on its own.
The evidence is one month: I pointed an orchestrator I had built at my own repositories and let it work unattended — planning, spawning agents, running gates, recording what happened. This site is the field guide that came out of it, with four longer pieces built from what the record actually showed. Some of it went well, a good deal of it did not, and both are published, because a number nobody can check is an opinion with a decimal point.
The ten concepts
In order, but there is no wrong door — each page states the idea in plain language first, and only then shows how one orchestrator implements it.
the orchestratorruns sessions back to back while nobody watches — picks the work, runs the gates, writes the record
the harnesswhat one session is handed — the window, the tools, the meter, what came before
the loopone session — the loop is the agent
the model is the one part you did not build — every other ring is yours
- Concept 1Agentic engineering
The engineering is not in the agent. It is in the loop around it.
- Concept 2Multi-agent orchestration
The plural is in the roles, not in the fleet.
- Concept 3Context engineering
Deciding what an agent is allowed to know, and paying for it.
- Concept 4Token economics and context budgets
A session's cost is not its length. It is roughly its length squared, and that changes every decision about it.
- Concept 5Evals, gates and acceptance
A gate is a command and an exit code. Everything else is an opinion with a schedule.
- Concept 6Independent verification
The thing that checks cannot be the thing that did it, and a prompt is not a separation.
- Concept 7Durable execution and resumability
If the work is longer than the thing doing it, the work has to live somewhere the worker does not.
- Concept 8Human in the loop
Oversight is not somebody watching. It is a place the machine stops, one that names what it is waiting for and exactly what clears it.
- Concept 9Agent observability
The question is not whether it is logging. It is whether you can ask something nobody planned for and get an answer.
- Concept 10Agent memory
Everything a session works out dies with it unless something writes it down, and everything written down is paid for by every session after it.
Four longer pieces
Each one carries at least one number nobody else publishes, and each is written from the same store the concepts cite.
- Article 1What an autonomous run actually costs
Nobody publishes the bill. This is the whole of one, across every run in the corpus, with the money that bought nothing left in.
- Article 2Never believe the agent
Verification has to be a separate program with real exit codes. The argument for that is not in the gates that passed. It is in the ones that did not, and in the few that never ran at all.
- Article 3The nudge that sat below the median
Every autonomous run puts a ceiling on what one session may spend, and almost every ceiling was picked by feel. The ledger can settle it. The awkward part is that the number you most need is the one a ceiling deletes.
- Article 4The ledger that lied
An autonomous run keeps a record of itself, and that record is what every later decision gets made from. This corpus holds a population of sessions whose record is empty in four places and full in the fifth. The shape of the hole turned out to be a code path, and the number it produced was the reassuring one.
Where to start
The site reads in any order, so these are doors rather than chapters — take the one that sounds like the question that brought you here.
new to all of this
What an agent actually is, and where the engineering genuinely lives — in the loop around the model, not inside it. The first concept assumes nothing and defines its words as it goes.
deciding what to trust
An agent reporting that its work is done is not the work being done. How checking gets separated from doing, and what a month of gate results said about believing the agent.
here for the bill
What a month of autonomous runs actually cost, where the money went, and why the length of a session dominates every other variable an orchestrator can turn.
The machine behind it
Every figure on this site was produced by one orchestrator — Conductor, a machine I built that plans, spawns agent sessions, gates them, and records what happened in a store the run cannot edit. What it is made of, how a run is declared, where each of the ten concepts lives in the mechanism, and the full record it kept — the corpus in numbers, the findings, the runs in depth — now live on their own page.
orchestrator · conductor — plan · limits · gates · record
Or start from a subject
The three sections say what kind of page something is. These say what it is about, and each one gathers concepts, articles and run reports together.
How to read this
No figure here was typed by hand. A page names an evidence key; the value comes from a corpus recomputed from Conductor's own run store, and a page citing a key the corpus does not have fails the build. The failures are published beside the wins, because the waste and the abandoned runs are part of what a run costs.