AI Coding — UltraCoding and agent harnesses

By the end of the course, you will be able to

• Turn a need into a decision-complete plan: scope, dependencies, acceptance criteria, exclusions and stop conditions.
• Describe the four harness subsystems and their responsibilities in a repository.
• Prepare focused context, least-privilege tool access, budgets and a single writing session.

And also

• Run a change in an isolated exercise environment and trace the inputs and checks used.
• Have the result challenged by an adversarial review, then checked with tests and audits suited to the case.
• Assemble an evidence pack, state residual risks and propose a verifiable recovery path.

Audience

Developers, tech leads, engineering managers, software quality specialists and delivery leads who want to govern coding agents working in an existing repository.

Prerequisites

Experience in software development, Git and a coding agent; ability to navigate a repository, run tests and interpret their results. Each participant brings a laptop and access to an organisation-approved coding agent.

States, observation, action, feedback, recovery and stop conditions. Distinguish a one-off answer from an acting cycle.

Available commands, files and connectors; contracts, least-privilege access and dependency changes.

Select useful rules, interfaces, documentation and history; bound the context and surface unknowns.

Exercise sandbox, reproducible tests, human approval, logs, limits and recovery procedures.

Scope, approve and bound execution

1. Opening and positioning (30 min) : Introduce the running exercise and establish the starting point.
2. Context and repository exploration (1 h) : Identify relevant files, tests, contracts, conventions and unknowns.
3. Workshop1: a decision-complete plan (1 h 30) : Explore without changing files; specify criteria and scope, then review and approve the plan.
4. Harness, permissions and one writer (1 h) : Set tool access, least privilege, budgets, stop rules, logging and a single writing session.
5. Workshop2: bounded implementation (1 h 30) : Implement the approved change in the training repository and retain an inspectable Git state.
6. Review and measurement (1 h 30) : Compare plan and result, record corrections, time, available costs and stop decisions.

Challenge, verify, recover and decide

1. Recap and formative quiz (30 min) : Revisit the first day’s decisions and risks.
2. Adversarial review (1 h) : Challenge assumptions, edge cases, excessive permissions and unexpected effects with a separate reviewer.
3. Three-layer verification (1 h 30) : Run behavioural tests, relevant security/accessibility/internationalisation checks, and inspect the observable result.
4. Workshop3: incidents and recovery (1 h 30) : Handle a stale plan, hostile instructions and a repeated operation; suspend, verify, recover and document.
5. Evidence pack and release decision (1 h) : Link the plan, Git state, checks and remaining gaps to the actual candidate; simulate approval and rollback.
6. Final assessment and transfer (1 h 30) : Present the evidence pack, take the quiz and define a personal transfer plan.

Running exercise

A supplied practice repository contains a regression defect and a documented incident. Participants establish a baseline, define the authorised change, make a bounded modification and test the implementation against reproducible checks.

Individual or pair deliverables

• Map of the four subsystems and their boundaries.
• Decision-complete plan with mandate, acceptance criteria and exclusions.
• Execution log and results of the checks actually performed.
• Evidence pack linked to the candidate, residual risks and recovery or rollback procedure.

Assessment of learning

Continuous assessment of deliverables; five practical criteria scored 0–2, with a 7/10 threshold and full scores required for authorisation and recovery. A ten-question final quiz also uses a 7/10 threshold. Participants present the actual candidate, its evidence and the limits of verification.

Follow-up

One-hour positioning interview and report, learning materials, standard library and training lab access for 90 days, and a 30-minute individual video coaching session at 90 days. Completion documents and end-of-course satisfaction assessment.

Transferable practices

Specification, human authorisation, bounded execution, verification separate from generation, logging and recovery.

Limits discussed

Results depend on tests, permissions, repository, context and chosen tools. An agent review alone is not independent proof.