Agentic development at Pico
Agentic development at Pico covers the way AI agents are designed, governed, and integrated into professional development and business environments. The focus is not on individual prompts or experiments, but on making AI agents reliable, repeatable, and organisationally useful actors. This requires discipline, clear frameworks, and a shared language for how agents work with tasks, data, and decisions.
In practice, Pico sees three interconnected disciplines as essential: harness engineering, context engineering, and feedback loops. Together, they turn agentic development into an engineering discipline that can be scaled and embedded in the organisation.
Harness engineering – control and predictability
Harness engineering is about defining the framework within which an AI agent operates. This is where the agent's role, responsibilities, and workflow are established. An agent does not become valuable because it can do a lot, but because it does the right thing in the right way – every time.
At Pico, harness engineering is understood as the work of: defining the scope and success criteria of a task, governing how the agent plans its work, establishing how execution happens step by step, describing how output is reviewed and approved, and determining whether and how the agent documents and updates status in existing tools and systems.
This framework is what distinguishes an experimental agent from one that can operate in a professional setup with accountability, traceability, and repeatable value. Harness engineering makes AI work predictable and applicable in larger teams and organisations.
Context engineering – relevant knowledge, nothing more
Context engineering is about ensuring that the agent has access to precisely the context needed to complete a task correctly. This includes both content and constraints. An agent that receives too little context will guess. An agent that receives too much becomes inefficient and imprecise.
In Pico's understanding, context engineering involves giving the agent: a clear and unambiguous task description, access to relevant documentation, standards, and workflows, insight into the technical stack, architecture, and conventions, and knowledge of domain-specific concepts and data – for example, directly from a PIM or ERP system.
At the same time, context engineering is about filtering. Rather than giving the agent all historical material, mechanisms such as search, structured access, and context layers are built so the agent only draws on the knowledge relevant to the specific task. This is a prerequisite for stable and reproducible results.
Feedback loops – from assumptions to knowledge
Once the agent is properly harnessed and has the right context, the next question arises: is the agent actually delivering the desired result, and does it improve over time? This is where feedback loops come in.
Feedback loops are the mechanisms that measure and evaluate the agent's output. These can be automated evaluations, human reviews, or analyses of error and success rates across task types. Without feedback, adjustments to harness and context are based on gut feeling rather than knowledge.
At Pico, feedback loops are seen as what makes agentic development measurable and improvable. They ensure that changes to the setup actually produce better results, and that experience can be shared and reused across teams.
How agentic development is organised in practice
The pattern Pico observes in many organisations is that a small number of people quickly achieve strong results with AI agents. The challenge arises when those experiences need to scale. The value often remains local, because the rest of the organisation lacks the frameworks and intuition to work at the same level.
This points towards a more specialised structure, where a smaller number of roles work dedicatedly on harness and context engineering. These roles build and maintain a shared platform for AI agents, tailored to the company's domain, processes, and system landscape. The platform ensures consistent context, standardised workflows, and built-in feedback mechanisms.
Developers and professional users can still experiment freely, but broad value creation only emerges when AI agents can be used without everyone needing to be experts in agent configuration at the same time.
The foundation for success
Agentic development starts earlier than the execution of a task itself. For AI agents to operate reliably, clear task descriptions, well-defined acceptance criteria, and up-to-date documentation are required. These elements are part of organisational discipline – not merely technical details.
When the platform is built correctly, AI agents can themselves contribute to improving these very foundations. This closes the loop: harness engineering and context engineering become not just technical tools, but a way of structuring work, responsibility, and knowledge across the organisation.