The Tekton Principles.
Every post advances one durable idea about building with agents. They are numbered, they do not get reused, and they accumulate here into a working canon. The list growing is the argument: the thinking compounds, the same way the work does.
8 principles and counting.
- ◆01
Design the workflow, not the output.
Treat agents like a delivery pipeline, not a vending machine. Spend the effort designing the deterministic workflow agents follow, not the one-off result. Like DevOps automation, a good workflow compounds: every reuse pays back in token efficiency and time-to-build.
From: How we built this site with Hekton → - ◆02
Trust infrastructure is the product, not the model.
The scarce resource in agentic engineering is not model capability. It is the operating model around the model: classification, decision logs, review gates, and a human-approval boundary. Build that first, and the model becomes a commodity you can safely swap out later.
From: Tools are easy. Operating models are hard. → - ◆03
Judgment doesn't disappear. It relocates.
Agentic delivery doesn't remove human judgment from engineering. It moves that judgment from doing the work line by line to defining the workflow the agent follows and deciding when to trust its output.
From: The roles don't disappear. They move. → - ◆04
The struggle is the point, not the output.
A rubric fix that only reduces spoilage for a human reader does nothing against a coding agent that already knows the answer. Say plainly what the fix protects and what it doesn't, then give the learner a real check for the gap it can't reach.
From: The Struggle Is the Point: What Agentic Harnesses Do to How We Learn → - ◆05
An uncalibrated judge is a vibe with a schema.
A judge that sounds confident has not been measured. Calibrate against known-correct answers first, and check the judge's ability to reject, not only its ability to agree.
From: LLM-as-Judge, and How Not to Fool Yourself → - ◆06
Context is infrastructure. Budget it like code.
The code is not what a future agent or engineer resumes from. The decision log is. Design it, review it, and prune it with the same rigor as the services it explains.
From: Context Is the Architecture → - ◆07
A rule that bends when the agent is confident isn't a rule.
Governance that only fires when the agent recognizes risk is the agent's own judgement wearing a policy label. Real governance binds even when the agent is sure it doesn't need to.
From: Governance That Survives the Agent's Own Confidence → - ◆08
Route on evidence, not on how expensive it feels.
The safest-feeling default, always the frontier model, is not the same as the correct one. Measure which tier a task actually needs before you pay for capability it doesn't.
From: Cost-Aware Model Selection: When Local Beats Frontier →