Module 04 · Agentic Engineering

Scope

The object of engineering is the complete trajectory of observations, intermediate decisions, tool calls, state changes and feedback. A model, its harness and the surrounding environment form the effective system. The harness supplies technical controls; responsibility for scholarly or organisational adequacy remains with the authorised role.

• Distinguish model, harness, agent, tools, state and environment by function.
• Bound a multi-step task through permissions, reversible increments, stopping conditions and observable checks.
• Integrate delegated evidence into maintained project knowledge and an explicit acceptance decision.

Concepts

Starter sequence

1. 1. Treat capability as a compound system

   An observed result depends on the model, the supplied context, the harness and the environment. Comparing models alone leaves tools, permissions, state and feedback outside the analysis.

2. 2. Engineer the action and feedback loop

   A controlled loop registers the current state, selects one bounded action, observes its consequence and revises the next step. The final artefact is only one part of the evidence.

3. 3. Set permissions and reversibility

   Tool access follows the least-privilege principle. Source material stays protected, generated outputs remain confined to an authorised workspace and publication actions require explicit release.

4. 4. Assign information to the right mechanism

   Persistent instruction files carry recurring project rules. Skills package reusable procedures. Model Context Protocol resources connect tools and data. Each mechanism has its own discovery, permission and maintenance boundary.

5. 5. Delegate bounded work

   A Subagent receives a defined question, source frame and return format. Parallel sessions and Worktrees isolate contexts and file changes. The parent trajectory integrates evidence and resolves disagreement.

6. 6. Preserve evidence through write-back

   Promptotyping links implementation to documented project knowledge, verification and acceptance. Research Mission Control adds explicit authority, bounded lanes and evidence-mapped integration.

Learning routes

• Available · local tutorial

  Claude Code introduction

  A product-specific route through persistent instructions, tools, Skills, Plugins, Subagents, Worktrees and visual verification.

• Available · local tutorial

  Codex introduction

  A product-specific route through installation, project instructions, Skills, delegated work, Worktrees and observable verification.

• Available · local tutorial

  Pi introduction

  A product-specific route that distinguishes the locally running harness from its model provider and inference location.

• Available · external method

  Research Mission Control

  The public repository documents accountable orchestration through Research Director, Research and Operations roles.

• Available · external method

  Promptotyping

  The method connects maintained project knowledge, agentic implementation, verification, acceptance and write-back.

Curated Skills

These externally maintained MIT-licensed examples show distinct procedural functions. Every entry is presented as a reference. Its complete instructions, side effects and permission scope still require a KCE review before any recommendation or installation.

• grill-with-docs

  Clarifies a plan through a structured interview while maintaining shared terminology and decision records.

  Invocation

  User-invoked

  Compatibility

  Claude Code plugin; Codex and compatible agents through editable skill files

  Write scope

  Project domain document and architecture decision records

  Review status

  Reference · instruction and permission review pending

  040203Source file

• research

  Investigates one bounded question against high-trust primary sources and returns a cited repository artefact.

  Invocation

  Model-invoked

  Compatibility

  Claude Code plugin; Codex and compatible agents through editable skill files

  Write scope

  One cited Markdown result at the task-authorised repository path

  Review status

  Reference · instruction and permission review pending

  040305Source file

• domain-modeling

  Tests project vocabulary against edge cases and updates the responsible domain and decision artefacts.

  Invocation

  Model-invoked

  Compatibility

  Claude Code plugin; Codex and compatible agents through editable skill files

  Write scope

  Project domain document and architecture decision records

  Review status

  Reference · instruction and permission review pending

  0304Source file

• diagnosing-bugs

  Builds an observable failing condition, minimises it, tests hypotheses and closes with a regression check.

  Invocation

  Model-invoked

  Compatibility

  Claude Code plugin; Codex and compatible agents through editable skill files

  Write scope

  Task-bounded instrumentation, fix and regression test

  Review status

  Reference · instruction and permission review pending

  0405Source file

• tdd

  Constrains a feature or fix through a red-green-refactor loop implemented in small vertical slices.

  Invocation

  Model-invoked

  Compatibility

  Claude Code plugin; Codex and compatible agents through editable skill files

  Write scope

  Task-bounded tests and implementation

  Review status

  Reference · instruction and permission review pending

  0405Source file

• code-review

  Compares a diff independently with repository standards and the originating specification.

  Invocation

  Model-invoked

  Compatibility

  Claude Code plugin; Codex and compatible agents through editable skill files

  Write scope

  Review evidence; implementation changes require a separate authorised task

  Review status

  Reference · instruction and permission review pending

  0405Source file

• handoff

  Compacts a working conversation into an explicit state transfer that another agent can continue.

  Invocation

  User-invoked

  Compatibility

  Claude Code plugin; Codex and compatible agents through editable skill files

  Write scope

  One handoff document at the configured project location

  Review status

  Reference · instruction and permission review pending

  0403Source file

• writing-for-agents

  Structures persistent instructions and skill documents for reliable discovery, retrieval and use.

  Invocation

  Model-invoked

  Compatibility

  Claude Code plugin; Codex and compatible agents through editable skill files

  Write scope

  Task-authorised agent instruction and skill documents

  Review status

  Reference · instruction and permission review pending

  0304Source file

External collection status: maintained outside KCE, MIT licence, checked 25 August 2026. The source repository supports a managed Claude Code plugin and editable installation for Codex and other compatible agents.