Topic hub
How to design bounded agent roles, contracts, state, and independent verification without replacing a simple pipeline with apparent autonomy.
Position
Agentic engineering here means designing a bounded work process rather than giving a model broad freedom to act. Each step needs an input artifact, permitted tools, a state record, and a completion condition. That makes it possible to check whether an agent changed the right file, used an appropriate source, and received approval for the next step. The material does not promote automation for its own sake: it also shows when a simple script, retrieval step, or manual review gives a safer result. Rollback and independent verification belong to the function, not to an after-incident appendix.
Practice starts with a small, reversible step: who starts an action, on which artifact, with which control, and where the result is recorded. Only after that description is it useful to decide whether a task belongs to an agent, a person, or a simple deterministic process.
An agent role needs bounded input, a named goal, allowed tools, an output artifact, and a completion condition. Without that, the word ‘agent’ hides an unclear process in which nobody can tell who did what. A good contract also makes it possible to ask whether a simple call, retrieval step, or deterministic routine solves the task without enlarging the error surface.
Material workflow state should be recorded in artifacts that can be read after execution: a specification, input file, decision record, patch, or test report. Conversation memory is not a sufficient source of truth for a process a team must take over. When the next step cannot identify its input and state version, it should not claim repeatable execution.
Independent review checks the product of work rather than merely repeating its description. It can be a test, diff inspection, command reproduction, or comparison with an acceptance criterion. The person or role that prepared a change should not be its only approving instance in a high-risk workflow. Separating execution from control creates a place to detect an error before it becomes the next automated action.
Automation should state what can safely be reversed, which artifact represents the previous state, and who decides to stop. Rollback is not a declaration of confidence or the final page of a runbook. It is a condition of an experiment, particularly when an agent can change files, call tools, or pass output to later participants in a workflow.
In practice
An agent role is useful only when it has a bounded contract, input state, named artifact, and completion condition. Independent verification is not a second account of the same work: it checks a file, test, PR, or result that the implementation author cannot approve alone.
RAG and agent-workflow design helps choose a simple pipeline before adding another role.
Evidence Path Protocol provides a minimum trace of specification, test, and acceptance decision.
Projects
Prototype / Agentic Engineering
An original course and protected environment where educators and researchers build an agent workspace through inspectable artifacts.
The result combines a public nine-session curriculum with a protected teaching space, a versioned material corpus, search, and course packages.
Author and instructor · Curriculum design · Workspace architecture · Project mentoring
live demo · live demo
Open source / Agentic Engineering
An open two-role workflow: specification → code → pull request → CI, with separated responsibilities and quality gates.
The repository demonstrates a controlled agent workflow rather than code generation in isolation.
Workflow design · Specification · Quality gates · Open-source publication
repository · article
Start here
An educational article about the path from Chat Completions through the Responses API and Agents SDK to managed agents, and the engineer’s changing responsibility.
The value of an agent workspace comes not from one answer but from a controlled project that preserves instructions, artifacts, tests, and a history of repeatable work.
Controlled agent work starts with a contract, role separation, and an external gate—not with the number of agents.
Public code can reveal process, tests, and decisions, but only scope and limitations make it interpretable as evidence.
Reading
An article about organising work with AI so that a result remains connected to sources, successive tasks, verification and human decision-making.
An educational article about the path from Chat Completions through the Responses API and Agents SDK to managed agents, and the engineer’s changing responsibility.
The value of an agent workspace comes not from one answer but from a controlled project that preserves instructions, artifacts, tests, and a history of repeatable work.
Controlled agent work starts with a contract, role separation, and an external gate—not with the number of agents.
Public code can reveal process, tests, and decisions, but only scope and limitations make it interpretable as evidence.