How we work

Our asset lies in the working method that allows us to adapt workflows and the most suitable agentic development processes for each client and project, taking advantage of technological evolutions. These are the 6 elements of the method of adesso.it: 

1. Spec-driven delivery: in each iteration, we define objectives, constraints, and acceptance criteria together with the client, written in a form that is readable by both people and agents. 
2. Context engineering: an agent works well only if it truly knows the client: their domain, their application architecture, their business rules, their regulatory constraints. We structure this knowledge in a form that can be queried by agents so that each iteration starts from the client’s real context, not from a generic model. 
3. Model routing and FinOps: there is no single right model for everything. In the planning and execution phases, we route each task to the most suitable model, by capability and cost, instead of always using the same engine for everything. The resulting control is not over the cost of a single inference but over the cost per outcome: the metric that makes the scalability of the project sustainable over time. 
4. Harness engineering: agents are powerful but fallible: verification is the phase on which we invest the most engineering and time. We build a harness of automated tests, guardrails, and human review checkpoints that intercept errors before they reach production. This is what makes reliable work that, by its nature, is born non-deterministic. 
5. Governance by design: permissions, data management, and traceability of decisions are designed from the outset, calibrated on the governance model, contract, and sovereignty requirements of each individual client. This is what allows a solution to move from a controlled environment to real use in production, without surprises along the way. 
6. Learning loop: each iteration leaves something in legacy for the next: skills, playbooks, reusable controls. We close the cycle by collecting what worked and what didn’t, bringing that learning into the following project. Thus, our working method improves with each iteration. 
   
   The working method is cyclical and iterative, divided into 7 phases: Intent, Context, Plan, Execute, Verify, Release, Learn. We can return to the previous phase at any time: it is our agile approach, based on continuous iteration cycles together with the client.

Agentic Coding produces acceleration on closed and verifiable specific activities autonomously by agents: writing and executing code, front-end and back-end testing, refactoring on known patterns, documentation generation. We call this type of activity execution layer. 

The same acceleration is not recorded in all activities requiring context: architectural decisions with cross-impact, human validation of knowledge extracted from legacy systems, coordination with external partners. We call this other type of activity situational layer. 

A real case documented by adesso.it, an enterprise project in the education sector on a brownfield basis, shows the difference in numbers: the infrastructure setup estimated in 30-40 man-days was completed in 11 man-days, with an acceleration of about 3 times, and over 900 test cases were written and executed from the start of the project, while two weeks of development remained stalled waiting for a decision on the external partner side, not due to a technical limit. The lesson is clear: gaining speed in the execution layer makes sense only if the situational layer directs it in the correct direction. Otherwise, the risk shifts further ahead, where correcting it costs more. Added to this is a second dimension of benefits: higher software quality because objectives, constraints, and verifications are explicit from the first line of code, full traceability of decisions even months later, reusable context on subsequent projects, and governance firmly in the hands of the people.

The distinction between the executive and situational layers explains how the weight of the five standard phases of a project changes when Agentic Coding comes into play. Previously, the five macro-activities maintained a relatively balanced weight; now the balance changes asymmetrically. 

• Strategy and Discovery belongs entirely to the situational layer and becomes even more crucial: in an era of abundant technical solutions, understanding what the priority problem to solve is counts more than knowing how to build it, and this choice cannot be delegated to agents.  
• Design and Concept remains in the situational layer and retains all its weight: without solid design, elegant solutions risk being centered on the wrong need.  
• Architecture Design lives on the threshold between the two layers: agents support the production of diagrams and the evaluation of alternatives, while decisions with cross-impact on the application landscape, security, and data model remain governed by people.  
• Development is the properly executive phase, where acceleration is more visible with an additional cost linked to AI tokens and quality and security standards that remain high.  
• Testing and Deploy also lives on the boundary: AI supports the generation and execution of tests well, but precisely because the code is generated by AI, validation requires an even more structured process, where human judgment on what to verify thoroughly remains decisive.