Agentic Development: Context as Core Competency
Agentic development represents a fundamental paradigm shift driven by non-determinism and intent-based workflows. This analysis explores how context management replaces traditional code-centric practices, introducing a Context Development Lifecycle (CDLC) that integrates with the SDLC. Learn how to mitigate LLM biases, manage costs, and establish continuous evaluation frameworks for scalable AI-driven software engineering.
The Paradigm Shift in Agentic Development
The emergence of agentic development marks a fundamental departure from traditional software engineering, driven by three core challenges: non-determinism, intent-centric workflows, and unprecedented velocity. Unlike deterministic builds, agentic systems require new operational frameworks to manage variability and ensure reliability. The primary strategic implication is that context, not code, has become the critical asset in software development. Organizations must transition from code-centric to spec-centric models, where the ability to articulate intent and constraints defines success.
Context as the Core Competency
Context management serves as the primary mechanism for aligning human intent with agent execution. This involves a structured lifecycle that includes defining, evaluating, distributing, and observing context. Just as DevOps transformed operations through continuous integration, the Context Development Lifecycle (CDLC) transforms development by treating context as a living artifact. This lifecycle requires continuous evaluation to detect context rot and ensure that agents operate with up-to-date, accurate information. The hierarchy of context, ranging from organizational policies to application-specific documentation, ensures that agents inherit necessary constraints while maintaining flexibility.
Operational Implications and Cost Management
Agentic development introduces significant operational complexities, including a bias toward re-implementation over reuse and the inherent obsolescence of LLM training data. These factors necessitate active management of agent behavior to prevent technical debt and security vulnerabilities. Furthermore, the cost of agentic workflows has shifted from a cost-saving measure to a major operational expense. Enterprises must implement robust cost management strategies, monitoring token usage and model selection to optimize for efficiency. The integration of agentic workflows into the existing SDLC requires new tools for evaluation and monitoring, ensuring that agents perform reliably in production environments.
Strategic Recommendations
To succeed in this new paradigm, organizations must invest in context engineering capabilities. This includes developing evaluation frameworks that mimic traditional testing hierarchies, from lightweight unit tests to comprehensive end-to-end scenarios. By treating context as a first-class citizen, companies can leverage the speed of agentic development while maintaining the quality and security standards required for enterprise software. The future of software development lies in the seamless integration of human intent and machine execution, facilitated by robust context management systems.
Key insights
-
Agentic development is non-deterministic, meaning repeatable processes are difficult to establish. This requires a shift from deterministic build verification to continuous monitoring and evaluation of agent behavior.
Impact: Organizations must invest in continuous evaluation frameworks to mitigate the risks of non-deterministic outputs, ensuring reliability in production environments.
-
Software development is shifting from code-centric to intent-centric, where code is disposable and specifications define the outcome. This change requires developers to focus on articulating constraints and goals rather than implementation details.
Impact: Teams that master intent-based specification will leverage agentic speed more effectively, reducing the time spent on manual code review and maintenance.
-
LLMs exhibit a bias toward re-implementing functionality rather than reusing existing libraries, leading to potential security and maintenance issues. This bias must be actively countered through explicit context rules that enforce reuse.
Impact: Failing to address this bias results in increased technical debt and security vulnerabilities, as re-implemented code lacks the benefits of centralized updates and patches.
-
Agentic development costs have shifted from a cost-saving tool to a significant operational expense, requiring active management of token usage and model selection. This mirrors the cost management challenges of cloud infrastructure.
Impact: Enterprises must implement cost monitoring and optimization strategies to prevent runaway agentic spend, ensuring that AI-driven development remains financially sustainable.
-
Context is a perishable asset that rots over time, requiring continuous maintenance and evaluation to remain accurate. This necessitates a Context Development Lifecycle (CDLC) that mirrors DevOps practices.
Impact: Implementing a CDLC ensures that agents operate with current, accurate information, preventing errors caused by outdated context and improving overall development efficiency.
Action items
-
Establish a Context Development Lifecycle (CDLC) that includes stages for defining, evaluating, distributing, and observing context. Integrate this lifecycle into the existing SDLC to ensure continuous improvement of agent performance.
Impact: A structured CDLC reduces the risk of context rot and ensures that agents consistently align with organizational goals and technical standards.
-
Implement hierarchical context management, separating company-wide policies, business unit constraints, and application-specific documentation. This structure allows agents to inherit global rules while adapting to local requirements.
Impact: Hierarchical context ensures consistent enforcement of security and design standards across the organization, reducing variability in agent behavior.
-
Develop evaluation frameworks that mimic traditional testing hierarchies, including lightweight unit tests and comprehensive end-to-end scenarios. Use these evaluations to monitor agent behavior and detect context failures.
Impact: Robust evaluation frameworks provide early warning of context issues, allowing teams to intervene before errors propagate into production systems.
-
Enforce reuse of existing libraries and functions through explicit context rules to counter LLM bias toward re-implementation. Regularly audit agent outputs to ensure compliance with reuse policies.
Impact: Enforcing reuse reduces technical debt and security vulnerabilities, ensuring that agents leverage the benefits of centralized code maintenance and updates.
-
Implement cost monitoring and optimization strategies for agentic workflows, tracking token usage and model selection. Adjust model configurations based on task complexity to balance speed, quality, and cost.
Impact: Active cost management prevents runaway agentic spend, ensuring that AI-driven development remains financially sustainable and aligned with business objectives.
Quotes
“The primary challenges in agentic development are, one is it's non-deterministic. The second is that it revolves around intent, not code. The third is that agentic development is faster. than anything we've seen before.”
“I do think that the hammer that we have is context. LLMs eventually are just stateless machines that we pass. a bunch of context to. They calculate their weights and they figure out what the next words are.”
“The core competency in agentic development is context management. And there are a lot of types of context. There are rules that you are kind of explicitly and aggressively pushing. There are skills that you are hinting and making available to the agent to try and pull down.”