Imad Touil is an engineering leader at QuantumBlack, AI by McKinsey, focused on how enterprises reorganize software delivery around AI agents. His distinctive contribution is an approach to agent skills as organizational knowledge: reusable instructions and executable procedures that preserve a company’s engineering practices, domain expertise, and operating rules across teams.
From building applications to shaping engineering organizations
His earlier work included web and mobile development at WebFullCircle in 2012, encompassing customer-management software, middleware, and business-process integration. At MyOpla, he developed a mass-mailing platform, taking responsibility for specifications, application architecture, and Scrum-based development.
He subsequently moved into engineering leadership at BN Corporation, associated with Cue Connect. His responsibilities expanded to distributed teams, technology roadmaps, product priorities, and the relationship between technical strategy and business goals. He worked on microservice architecture and continuous delivery on AWS, while coordinating development, operations, and quality assurance. Alongside that work, a 2017 consulting engagement at INWI involved agile coaching, workflow improvement, and operational reporting. These earlier engineering roles involved both building software and changing how organizations delivered it.
At QuantumBlack, Touil applies that combination to agentic software engineering. He introduced himself as a distinguished engineer in 2026, with a focus on making AI-assisted work usable across an organization rather than confined to individual developers. His argument about software delivery extends beyond implementation: product strategy, customer discovery, data preparation, infrastructure provisioning, launch, and incident response all shape whether software produces business value.
Making institutional knowledge executable
Touil’s approach to scaling agent skills borrows from established software architecture. A skill should do a specific job, remain reusable across workflows, and be discoverable by engineers who did not create it. He favors specialized, composable skills over large instruction bundles whose responsibilities overlap or conflict. Progressive disclosure supplies relevant instructions when needed, helping keep unnecessary material out of the agent’s context.
A regulatory-review workflow makes the idea concrete. Data-retention rules, disclosure standards, privacy requirements, and filing templates can become separate skills that a workflow combines. The intended result is consistent application of organizational rules across web, mobile, and other products, with review findings and audit artifacts that teams can inspect and act on. This is a proposed architecture for dependable execution, rather than a claim that instructions alone guarantee compliance.
Several concerns define his approach:
A skills registry needs human ownership. Touil proposes a searchable catalog containing metadata, dependencies, versions, and access controls. Engineers and agents should be able to find an existing capability before creating another one. Architecture, engineering, infrastructure, and security leaders must own the relevant domains and decide how changes are maintained.
Ungoverned skills become technical debt. Teams working with similar technologies can independently recreate the same instructions, accumulate conflicting versions, and lose track of who maintains them. Skills also require reevaluation as models change. Security review matters because imported skills can contain prompt injection or executable scripts, while sensitive business logic may require restricted access.
Govern the whole workflow. A well-maintained skill cannot compensate for a poorly designed delivery process. Touil extends the catalog idea to reusable workflows that include the necessary skills and supporting components. Teams should be able to retrieve, run, test, and improve those workflows, then contribute useful changes back to the organization.
Automatic evolution increases the need for controls. He treats static checks of skill structure and invocation as a useful starting point for evaluation, while recognizing that the broader evaluation problem remains unsettled. An automated improvement loop can accelerate changes without resolving ownership, security, or quality. His team simulation illustrates the intended benefits of shared governance; it does not establish measured enterprise productivity gains.
Touil’s writing on adopting agentic engineering gives this platform-oriented thinking a practical sequence. Start with a small group using one platform on noncritical work, observe the failures, and encode architecture conventions, security rules, domain language, and testing standards. Build custom infrastructure when actual constraints justify it. He argues that faster code generation shifts attention toward intent, system design, and quality governance—and warns that vague instructions and unmonitored agents can accelerate the accumulation of poor software. For Touil, the lasting investment is the organization’s ability to make its knowledge usable, maintainable, and shared.
Imad Touil explains how reusable agent procedures become organizational infrastructure: specialized skills, shared catalogs, model-aware testing, and human owners who keep the whole system useful as it grows.
The coding-agent sequence of specify, plan, task, and implement builds a product increment inside a larger lifecycle of strategy, discovery, data preparation, operations, and improvement.
Specialized, discoverable skills let teams reuse procedures and compose them into workflows. Progressive disclosure reduces context cost by loading relevant knowledge when needed.
A shared registry needs metadata, dependencies, versions, access control, evaluation, and observability—and named people responsible for maintaining its domains.
Retest skills against their tasks and changing models. Automatic evolution increases the need for governance because it accelerates changes to shared procedures.
▶ Watch ↗