What does directed agentic engineering look like from the client side?

Clients see a delivery team that moves faster than a traditional team, with clearer documentation, explicit decisions, and review gates at the points that matter. The pace changes; the visibility increases.

What changes for the client compared with a traditional engagement?

The cadence is shorter, the artefacts are richer, and the client is asked to make architectural decisions earlier. In return, surprises late in delivery drop sharply.

Insights

Whitepaper

Directed Agentic Engineering

The full case for Directed Agentic Engineering — what it is, why it matters, and how organisations can use it to become Solution Providers.

By: Consid AI Delivery Framework — early adopter edition
Published: January 2026

The core idea

A fork in how delivery works.

Tool User	Solution Provider
Adopts AI within existing model — faster developers, same processes.	Restructures coordination, knowledge accumulation, and governance with AI as core operating principle.
Output increases. Value proposition does not change.	Unit of value shifts from hours to outcomes, methodology, and accumulated insight.
Efficiency advantage erodes as every competitor does the same.	Differentiation compounds with every engagement.

Directed Agentic Engineering is the discipline of becoming a solution provider — not by using AI less, but by ensuring that human direction governs how AI operates.

Five principles

The operating conditions for structured delivery.

1
Think Before You Act
Before an AI agent touches a line of code, the intent must be clear. Speed without direction is not productivity. It is just more output.
2
Make Intention Explicit
Thinking that is not written down is invisible — to AI and to the team. Specifications and architecture make intention concrete enough to build against.
3
Make Knowledge Operational
Experience that lives in someone's head disappears when they leave. Procedures, workflows, and decisions must be codified as structured documents the whole team can follow.
4
Humans Take Critical Decisions
Human approval gates at the right points keep AI-accelerated delivery aligned with intent.
5
Execute in Small, Reversible Steps
AI agents can produce thousands of lines of code in minutes. Small steps are the discipline that keeps that power governed.

Three delivery modes

Not tool categories — operating models.

M1 — Assisted

Human leads, AI assists reactively. The developer writes; AI completes and suggests. AI is a faster keyboard. The human owns every decision and every commit. Where most teams start today.

M2 — Augmented

Human and AI in active collaboration. AI generates from specification; the developer reviews and owns the result. The developer stops writing code from scratch and starts writing specifications instead. Where the five principles become fully operational.

M3 — Autonomous

AI agents drive, human orchestrates and supervises. Multiple agents build in parallel. A human Supervisor sets policy, monitors output, handles exceptions, and approves outcomes. The direction the industry is heading.

What M2 looks like

Four things that change in every project.

1
Specification before generation
Developer writes a structured specification; AI generates the implementation — developer reads every line but does not write it from scratch.
2
Architecture before code
Architecture is established before a single line is generated and kept current throughout. The Design phase is the formal gate before development begins.
3
Attribution on every commit
Every AI-assisted commit carries a trailer recording which tool was involved. The provenance of every line of code is auditable.
4
Human approval technically enforced
A CI gate enforces that no code merges without a named human approver. A technical constraint, not a process recommendation.

Governance

What unreviewed AI output costs.

The efficiency gains M2 delivers are only worth capturing if the output is trustworthy. Speed without governance is how AI-assisted delivery produces the opposite of its intended result.

The Four-Step Review
Understand, Verify, Challenge, Decide. Every AI-generated output goes through this. Applies to code, architecture, test cases, and documentation.
CI Quality Gates
P0 and P1 findings block the merge. No exceptions. The pipeline enforces the human review requirement as a technical constraint.
Attribution
Every AI-assisted commit carries a structured trailer identifying the tool. The codebase is auditable for quality reviews, client audits, and regulatory enquiries.

The Delivery Workbench

The harness, pre-built.

Every engagement that runs M2 delivery needs a harness: the CI pipeline, agent configuration, specification templates, review standards, attribution policy, and governance rules. The Workbench is that harness, assembled and ready to adapt to context.

Data Model
Business capabilities, arc42 architecture, ADRs, specifications, and the traceability chain — the context agents need to act purposefully.
Playbooks
Structured workflows governing how work progresses. Each defines a trigger, entry criteria, steps, human gates, and exit criteria. Accumulated practice in a form the client owns.
Agents
AI agents configured to operate within the engagement's data model and playbooks. Planner, Executor, Verifier, Code Review Agent, Debugger, and Project Researcher.
Human Validation
Structured gates where human judgment is required before the workflow proceeds. Designed into the playbooks — not optional reviews added at the end.

Questions

Frequently asked

What does directed agentic engineering look like from the client side?: Clients see a delivery team that moves faster than a traditional team, with clearer documentation, explicit decisions, and review gates at the points that matter. The pace changes; the visibility increases.
What changes for the client compared with a traditional engagement?: The cadence is shorter, the artefacts are richer, and the client is asked to make architectural decisions earlier. In return, surprises late in delivery drop sharply.

Architecture before the first sprint. The foundation compounds.

Three things before the first sprint: architecture before code, CI pipeline configured, specification discipline agreed. From week three, the team is in M2 delivery.

Talk to Consid