A New Way of Working: Managing software delivery using Agentic Engineering

Development velocity has reached unprecedented levels, but it’s a metric that masks a deeper structural vulnerability. Delivery teams can convert tickets into code faster than ever, yet they struggle to verify whether the underlying requirements were unambiguous, whether the documentation reflects the live architecture and which party approved the decision encoded in the code. What constrains software delivery is no longer engineering throughput, but the quality of the requirements, documentation, tests and context on which the work is founded. Agents don’t remove that constraint. They expose it at greater volume and velocity than delivery organisations have previously had to absorb.

Familiarity is no longer the restraint. Wharton records 77% of US enterprise decision-makers as at least somewhat familiar with generative AI, while daily use in IT stands at 68%, a rise of 28 percentage points year on year.

Agentic deployment is the next phase. Here, the data separate stated intent from operational reality. McKinsey records 62% of organisations which are at least experimenting with AI agents and 23% that are scaling them somewhere in their business; however, a scant 10% report scaling agents in any individual business function. In software engineering, 77% report no agent use at all and 5% report scaling. The market reality is a stark dichotomy between widespread experimentation and negligible mature scaling. Neither budget nor model quality is the differentiator, but operational process.

Compression doesn’t eliminate process defects. It relocates them. Fragmented domain knowledge and obsolete documentation don’t impede an agent. They simply ensure that it works confidently to achieve the wrong objective. The result is highly plausible output that fails to meet business needs, violates architectural standards and accelerates technical debt.

The value of the analyst, the product owner, the architect and the domain expert increases as a consequence. Their function is no longer to transfer a task to the technical team, but to prepare the material that constitutes the basis for people and agents to operate on, formulating a single reliable understanding of the problem. A New Way of Working details this material as a prerequisite before entrusting an agent with a task at all.

Agentic engineering is the design and operation of governed AI agents that perform multi-step business tasks autonomously, integrating reasoning, planning and action with enterprise tools, data and workflows.

Accelerated delivery is the first observable benefit and it’s a genuine one. However, it isn’t an independent measure of success. An agentic environment amplifies the mechanism that has always produced technical debt. Agents generate substantial sections of a solution within minutes, but no volume of generated code can ever be a substitute for determining its comprehensibility, modularity, testability and capacity for extension by another team eighteen months later.

There is a second cost that attracts less attention. A large, loosely specified task assigned to a single model requires a broad context window, increases token expenditure and complicates oversight. A New Way of Working not only sets out why decomposition into smaller, verifiable stages is as much an economic decision as an engineering one, but also looks at how it alters the point at which errors are detected.

Contrary to prevailing market enthusiasm, executive confidence in unsupervised AI is receding. Capgemini records 22% of organisations expressing trust in fully autonomous AI agents, down from 43% twelve months earlier, while 74% consider that the benefits of adding human oversight to agent-driven tasks will outweigh the costs.

These two findings aren’t conflicting. They describe an organisational learning curve where, as operational experience accumulates, organisations specify the boundary of machine authority with greater precision. This is most acute when a decision carries financial exposure, health and safety implications or regulatory consequences. McKinsey's data points the same way. Of all the management practices tested, the greatest divide between the highest-performing organisations and the rest was recorded for defined processes determining how and when model outputs require human validation, with 65% of AI high performers reporting it, against 23% of the remainder.

The human-in-the-loop model is widely misread as manual approval of every minor action. A New Way of Working advances the opposite position, whereby the objective is the deliberate design of control points, with low-risk activity automated, high-impact decisions reserved for human judgement and auditability preserved end to end.

Agentic engineering restructures teams as much as it does processes. Roles centred on agent orchestration, context engineering, architecture and quality assurance increase in weight and, since the context being supplied now directly determines agent output, business roles change in parallel.

A more consequential problem arises further down the line. Deloitte not only records that entry-level work in the form of data entry and first-level customer support is being prioritised for automation, but also notes both that these roles are frequently the starting points for extended careers and that organisations will need to construct alternative pathways for professional advancement. Right now, that pathway is closing while the sector focuses on other matters.

An organisation’s readiness for agentic engineering begins not with tool procurement, but with an assessment of its current delivery process. A New Way of Working defines eight conditions as integral components of delivery. Three of them are:

●  a defined SDLC process with documented stages, owners and control points;

●  automated tests and CI/CD pipelines capable of detecting errors in AI-generated code and not solely in code authored by humans;

●  context, agent access, token consumption and operational cost monitoring as a continuous discipline rather than retrospective reconciliation.

The other five address requirements; documentation; agent-versus-human division of tasks; development of skill sets; and governance, security and auditability. Read as a set, these eight conditions function as a diagnostic. An organisation that meets them can let agents raise delivery pace without generating disorder. On the other hand, an organisation that falls short will end up identifying the missing conditions at production velocity.

A New Way of Working provides a solid framework. The more valuable conversation, though, is the one that turns the full focus on you and probes your delivery process.

Book a dedicated working session with MakoLab’s delivery team. They will map the report’s readiness conditions directly onto your organisation’s way of working and identify which gaps expose you to the greatest risk as delivery speed increases.

A New Way of Working: Managing Software Delivery Using Agentic Engineering is the second publication in a research series by Antal in partnership with MakoLab. It was preceded by AI-Ready Infrastructure: Observability and Operations.

Designed for IT leaders and decision-makers, the programme maps the operational reality behind enterprise AI adoption. Forthcoming reports will cover:

●   AI governance and data sovereignty;

●   the practical applications of neuro-symbolic AI

●   AI adoption and readiness across European enterprises; a comprehensive quantitative study tracking the current state of play.