Building for the Agent Era: Delegation, Context and Evaluation in AI-Native Development
Written by Rebecca Uffindell Fri 20 Mar 2026
At Tech Show London, Guy Podjarny approached AI agents not as a distant possibility but as operational infrastructure.
“Agents are no longer a theory. They are a reality.”
Drawing on experience founding Snyk and now leading Tessl, he described the transition from AI-assisted coding to agentic development as structural rather than incremental. The focus shifts from whether AI can support developers to how organisations learn to delegate to it effectively.
Delegation as a Development Model
Podjarny traced the evolution of AI tooling. Early coding assistants centred on augmentation, completing functions, suggesting syntax and accelerating execution. The developer remained firmly in control of implementation.
Agentic development changes that relationship. Developers increasingly articulate intent, define constraints, and review outcomes rather than writing every line of code directly.
“Conceptually, a better way to think about an agent is that it’s an AI employee, it’s AI labour that you are delegating work to.”
This reframing moves AI from feature to workforce. Delegation, however, introduces new managerial challenges.
Managing Non-Determinism
Unlike traditional software systems, AI agents operate probabilistically.
“It’s non-deterministic.”
Repeated queries can yield different results. A security check may detect a vulnerability in one run and miss it in another. Outputs vary, even when inputs remain constant.
In engineering cultures accustomed to deterministic systems, this variability introduces discomfort. Podjarny argued that organisations must build new competencies around measurement and evaluation.
“You cannot optimise what you can’t measure.”
Systematic output testing, regression detection, and structured feedback loops become foundational. Delegation without evaluation leads to inconsistency; delegation supported by oversight enables scale.
Podjarny compared this transition to earlier shifts in cloud and DevOps. As deployment speed increased, governance models evolved. AI now accelerates that pressure further.
“AI is moving faster than Cloud ever did.”
Controls designed for slower cycles require reassessment in AI-native workflows.
Context as a Primary Artefact
Podjarny also examined how the locus of engineering work is shifting. Historically, code served as the durable artefact of development. Requirements documentation often faded, and institutional knowledge filled the gaps.
In agentic workflows, context becomes central. The instructions, constraints, and preferences provided to an agent shape output quality more than manual code composition.
Code is easier to generate; precision in description becomes more consequential.
Context engineering involves calibrating scope, constraints and guidance. Excessive instruction can overwhelm; insufficient context can misdirect. The discipline lies in iteration and refinement.
Evaluation as Infrastructure
Podjarny introduced the idea of a “context development lifecycle,” suggesting that context should be treated with the same rigour as code: defined, evaluated, iterated, and deployed.
Evaluation requires organisations to define performance criteria, test agent outputs against those criteria, monitor behaviour in live workflows and adapt context as systems evolve. This process is ongoing rather than episodic.
Without structured evaluation, review fatigue and mistrust accumulate. With it, delegation becomes scalable.
Shifting Human Leverage
As agents assume more responsibility across the software development lifecycle, human roles adjust accordingly. Developers increasingly design policies, manage context and oversee evaluation frameworks rather than inspecting every line of generated code.
This represents a move from execution to orchestration.
New competencies emerge alongside this shift, and new tooling becomes necessary. Just as cloud adoption required DevOps practices to integrate security and deployment, AI-native workflows require governance systems built around delegation and oversight.
Cultural Adaptation and Experimentation
Productivity gains were discussed throughout the session, but Podjarny also emphasised the cultural dimension of adoption.
“Using agents doesn’t just breed efficiency. It brings Affinity.”
Hands-on experimentation shapes intuition. Organisations that avoid engaging directly with agentic tools struggle to define what effective delegation looks like in practice. Iteration builds fluency, and fluency builds confidence.
Redefining Engineering Leverage
Podjarny drew parallels to the cloud transition, when organisations that embraced new operational models accelerated ahead of those that hesitated.
Agentic development offers increased leverage, but leverage depends on structure. Delegation requires defined evaluation frameworks, calibrated context and disciplined governance.
The transformation underway does not eliminate the role of the developer, it reshapes it.
In AI-native environments, engineering leverage increasingly depends on designing the conditions under which code is generated, validated, and trusted.
Written by Rebecca Uffindell Fri 20 Mar 2026