Agentic development is the discipline of making intent, organisational context, and boundaries explicit enough for agents to execute safely, and outcomes verifiable enough for the system to decide when the work is done.

We value

  • Verified outcomes over generated output
  • Explicit intent and governed context over plausible interpretation
  • Executable guardrails over written expectations
  • Evidence-driven loops over one-shot delegation
  • Human judgment over human rubber-stamping

The items on the right still have value. Trustworthy delegation depends on the items on the left.

Our principles

1. Start with a clear task contract

Before delegating work, define the intended outcome, scope, non-goals, constraints, completion criteria, and escalation conditions. If any of these are materially unclear, the agent should ask rather than infer.

2. Manage context as organisational infrastructure

Context should not be recreated in every prompt or trapped inside individual projects. Organisational intent, policies, architecture, standards, domain knowledge, previous decisions, ownership, and approved exceptions should form a shared and governed body of knowledge.

Every piece of context needs a defined scope, authority, owner, provenance, and validity period. Agents should receive a resolved view of the context relevant to their task, with precedence, inheritance, exceptions, and conflicts made explicit.

3. Separate execution from verification

The agent may perform the work, but it must not be the sole judge of its success. Independent signals must determine whether the expected behaviour, constraints, and intended outcome have been satisfied.

4. No completion without evidence

"Done" is not a statement from the agent. It is a collection of verifiable results mapped to the expected outcome, including checks performed, observations made, assumptions taken, and risks remaining.

5. Make guardrails constrain, not merely advise

A useful guardrail defines what the agent may do next: continue, repair through an approved path, stop, or escalate. The environment should enforce critical boundaries instead of relying on instructions alone.

6. Do not let agents improvise across decision boundaries

Agents may choose implementation details within an approved space. They must not invent requirements, policies, business meaning, architectural decisions, or exceptions. Missing authority, stale context, contradictory sources, or material ambiguity must stop the loop and trigger escalation.

7. Grant bounded and reversible autonomy

Scope, permissions, time, cost, environments, and blast radius should be proportionate to the task. Prefer small, observable, and reversible changes over broad, irreversible action.

8. Put humans where judgment changes the outcome

Humans should own intent, trade-offs, exceptions, and consequential decisions. They should design the loop and resolve ambiguity instead of approving routine actions by default.

9. Turn every failure into system improvement

Repeated corrections should become better context, contracts, tests, guardrails, tools, or standards. Autonomy should expand only when evidence shows that the system can support it safely.

10. Measure the flow of safe outcomes

Measure time to verified value, rework, escaped defects, operational impact, and meaningful human intervention. Prompts, generated code, agent activity, and pull-request volume measure motion, not whether the work produced a safe outcome.

Before you delegate

An organisation should always be able to answer:

  • What outcome is expected?
  • What context applies?
  • Who owns that context?
  • What may the agent decide?
  • What must it never invent?
  • What evidence proves success?
  • What makes it stop?
  • Who owns the result?

The organisation provides the context.

The agent creates motion.

The system turns motion into evidence.

The standard turns evidence into trust.