Insights

The new engineering ethos: Why context is the product in AI-native software engineering

Sai Gade, General Manager – DevOps, SRE & Platform Engineering at UST

AI agents are only as effective as the context they can access. As software engineering becomes increasingly autonomous, success depends less on model capability and more on context engineering: making architectural decisions, business priorities, constraints, governance, and institutional knowledge machine-readable and available throughout the SDLC. Organizations that invest in engineered context, combined with human judgment and oversight, will be best positioned to achieve reliable, scalable AI-native software delivery.

Seventy years ago, two researchers at the Tavistock Institute in London, Eric Trist and Ken Bamforth, observed something unexpected in British coal mines.

New mechanized equipment had been introduced to increase productivity. On paper, the technology should have transformed productivity. Instead, output declined, morale suffered, and teams that had worked effectively for years began to struggle.

The problem was not the technology itself. The machinery had been introduced without considering how miners collaborated, shared knowledge, and organized their work. While the technology improved individual tasks, it disrupted the broader system that made the teams effective. Researchers later described this as a sociotechnical failure.

The lesson remains relevant today. As AI agents become embedded across the software development lifecycle (SDLC), many organizations are discovering a similar challenge. AI can generate code, automate testing, accelerate deployments, and support decision-making at unprecedented speed. Yet many failures attributed to AI are not technology failures at all. They are failures of context engineering.

Success in AI-native software engineering requires more than capable agents. It depends on the quality of the context available to them. That context shapes what agents know, the decisions they make, and the outcomes they produce. Managing it effectively is what separates fast agents from effective ones. It is a prerequisite for autonomous software delivery.

DIVIDER

Microsoft IQ validated the need for context

At Microsoft Build 2026, Microsoft and GitHub introduced Microsoft IQ, a context layer designed to help AI agents access and use enterprise knowledge. Capabilities such as Work IQ, Fabric IQ, Web IQ, and Frontier Tuning are designed to provide agents with greater awareness of business information, processes, and organizational context.

Microsoft IQ reflects a growing recognition that AI agents require more than powerful models to operate effectively. They need access to reliable context.

The Microsoft IQ context layer addresses part of that challenge by connecting agents to enterprise knowledge. However, software delivery requires more than enterprise knowledge.

Software systems are shaped by architectural decisions, dependencies, operational requirements, regulatory obligations, stakeholder priorities, and years of accumulated institutional knowledge. AI agents cannot reliably understand those factors unless that information is made available to them.

DIVIDER

The SDLC Context Planes: A framework for autonomous software delivery

Software delivery depends on multiple layers of context that influence how systems are built, maintained, governed, and evolved. The SDLC Context Planes framework makes those layers explicit.

The framework extends sociotechnical systems thinking into the era of AI-native software engineering. It recognizes that software delivery operates within multiple context planes that shape decisions, outcomes, and system behavior. These planes explain not only what a system does, but why it behaves the way it does, how it reached its current state, and where it is expected to go next.

The framework consists of three categories, 15 SDLC context planes, and a judgment meta-plane that provides human oversight when context alone cannot resolve competing priorities or outcomes.

Provenance: What the system carries

Provenance captures what a system carries about itself, its history, and the people who have shaped it over time.

• System plane: Provides context about the architecture, components, and environment in which a system operates.
• Dependency plane: Captures the relationships that influence how changes propagate across applications, services, and infrastructure.
• Boundary plane: Defines what a system is designed to do and the limits of its responsibility.

History matters as much as structure.

• Temporal plane: Provides insight into when decisions were made and the constraints that existed at the time.
• Decision plane: Captures the reasoning behind architectural choices, including alternatives that were considered and rejected.

Some of the most valuable context exists outside formal documentation.

• Cognitive plane: Represents the institutional knowledge often held by experienced engineers.
• Stakeholder plane: Reflects business priorities, user needs, and organizational expectations that continue to shape how systems evolve.

Constraint: What the system is bound by

Software delivery works within a set of constraints that influence every technical and operational decision.

• Security plane: Captures threat exposure, access controls, and risk considerations.
• Resilience plane: Focuses on recovery, continuity, and behavior under failure conditions.
• Operational plane: Reflects how systems function in production and what distinguishes normal behavior from abnormal behavior.

Organizations also function within established rules and obligations.

• Policy plane: Encompasses internal governance requirements, standards, and ways of working.
• Regulatory plane: Introduces external requirements, including legal, industry, and compliance obligations that cannot be ignored.

Direction: Where the system is going

Software systems are constantly changing. New requirements emerge, priorities evolve, and business objectives shift over time.

• Intentional plane: Captures the desired future state of a system, including product strategy, architectural direction, and long-term objectives.
• Governance plane: Defines accountability, decision rights, and the guardrails that guide change.
• Feedback plane: Provides signals about performance, outcomes, and effectiveness.
• Learning plane: Captures how knowledge evolves as systems, teams, and organizations adapt to new information and experience.

DIVIDER

The judgment meta-plane

Above all 15 context planes sits the judgment meta-plane.

Not every decision can be resolved through data, policy, or automation. Context planes may conflict. Business priorities often compete with technical realities. Regulatory requirements may constrain the best course of action. Some situations involve ambiguity, uncertainty, or consequences that cannot be easily reversed.

These are the moments where human judgment remains essential.

The judgment meta-plane serves as the human authority layer for autonomous software delivery. It provides oversight when competing forms of context cannot be reconciled and helps ensure that critical decisions remain accountable, explainable, and aligned with organizational objectives.

As AI systems become more capable, the importance of this layer only increases. Autonomous delivery does not eliminate the need for human decision-making. It changes where and how that judgment is applied.

DIVIDER

Why orchestration is not enough

GitHub Agent HQ’s arbitration capability reflects the principles of human-in-the-loop AI engineering, recognizing that AI agents will encounter situations where human oversight remains necessary.

In many ways, this aligns with the judgment meta-plane. When competing priorities, conflicting requirements, or uncertain outcomes arise, escalation to a human decision-maker is often the right response.

However, arbitration without sufficient context creates a different problem.

An agent that reaches a conflict point without understanding the decisions, dependencies, operational realities, or regulatory obligations surrounding a system can only surface the conflicts it can see. If critical context is missing, the agent may not recognize the most important issue.

This is why orchestration and governance are only part of the solution. Escalation mechanisms determine what happens after a conflict is identified. They do not address what an agent knows before it acts.

The gap is not the orchestration layer. It is the context that sits upstream of it.

Before an agent can make a recommendation, execute a task, or decide whether human intervention is warranted, it must have access to the context needed to understand the environment it is working within.

A well-designed context layer provides agents with machine-readable information about systems, decisions, constraints, dependencies, and objectives before execution begins. It helps ground actions in the realities of the software delivery environment rather than relying solely on model inference or incomplete information.

DIVIDER

Humans do not disappear; they migrate

A key implication of autonomous software delivery is that humans do not disappear from the software development lifecycle. Their role changes.

As AI agents assume more execution-oriented activities, including code generation, testing, analysis, and deployment support, human responsibility moves toward the areas where context, interpretation, and judgment remain essential. This is not a limitation of AI systems. It is a design principle.

The three human planes within the SDLC Context Planes framework—the cognitive plane, stakeholder plane, and judgment meta-plane—define responsibilities that should remain under human ownership.

DIVIDER

How UST PACE operationalizes context engineering

The SDLC Context Planes framework is intended to make context explicit, accessible, and actionable. UST PACE applies these principles to operationalize context engineering for AI-native software delivery.

PACE Lumen applies the principles of context engineering through a machine-readable layer that incorporates all 15 context planes and preserves awareness as systems, decisions, and business priorities evolve.

The goal is to provide agents with timely, relevant information that improves their understanding of the systems they support.

Organizations making the greatest progress toward autonomous software delivery are investing in engineered context, not just more capable agents.

DIVIDER

The future of software engineering: Context as the product

The lesson from the Tavistock researchers is still relevant today. The challenge was never the technology itself. It was understanding the system in which the technology ran.

The same principle applies to AI-native software engineering. The engineering ethos has not changed. Great engineers have always understood the systems they were building within. They preserved knowledge, documented decisions, respected constraints, and considered the people who would inherit and support those systems in the future.

Context engineering makes that ethos explicit, structural, and machine-readable.

AI models will continue to improve. Agent capabilities will expand. New tools, frameworks, and orchestration layers will evolve. Yet the effectiveness of those systems will depend on the quality of the context that surrounds them.

Organizations that invest in context will support reliable, accountable, and scalable autonomous software delivery. Those who do not may find themselves relying on agents that are powerful, but insufficiently informed.

The question is not whether your agents are capable. It is whether the context they rely on is engineered or archaeological.

Explore how UST and GitHub are helping organizations operationalize context engineering for AI-native software delivery.

DIVIDER

FAQs

What is context engineering in AI?

Context engineering is the practice of capturing, organizing, maintaining, and delivering the information AI systems need to function effectively. In software delivery, this includes business objectives, architectural decisions, dependencies, constraints, governance requirements, and institutional knowledge that help agents make informed decisions.

Why do AI agents fail in software delivery?

AI agents often produce unintended outcomes when they lack sufficient context. While agents can analyze code and automate tasks, they may not have visibility into historical decisions, business priorities, operational realities, regulatory requirements, and stakeholder expectations. As a result, technically correct actions may not align with organizational objectives or operational requirements.

What is a sociotechnical system in software engineering?

A sociotechnical system recognizes that software delivery involves technical components and human factors. Applications, infrastructure, processes, teams, stakeholders, and organizational dynamics all influence outcomes. Effective engineering requires understanding the technical system and the human system surrounding it.

How does context impact AI-driven development?

Context influences what AI systems know, how they interpret information, and the actions they take. Rich, accurate context helps agents produce more reliable recommendations, identify risks, understand constraints, and align actions with business objectives.

What are SDLC context planes?

SDLC Context Planes are a framework for organizing the information required for autonomous software delivery. The framework includes provenance, constraint, and direction planes, along with a judgment meta-plane that supports human oversight when competing priorities, ambiguity, or irreversible consequences require human judgment.

What is human-in-the-loop AI engineering?

Human-in-the-loop AI engineering is an approach that combines AI automation with human oversight. AI systems perform execution-oriented tasks, while humans provide institutional knowledge, stakeholder interpretation, governance, and judgment when context is incomplete or decisions have significant consequences.

How does Microsoft IQ work?

Microsoft IQ is a context layer designed to connect AI agents with enterprise knowledge. Through capabilities such as Work IQ, Fabric IQ, Web IQ, and Frontier Tuning, it helps agents access organizational information, business processes, and enterprise context to improve decision-making.

What is autonomous software delivery?

Autonomous software delivery refers to software development and operations processes where AI agents perform tasks such as coding, testing, analysis, deployment, and incident response with varying levels of autonomy. Success depends on effective governance, human oversight, and access to high-quality context.

DIVIDER

Resources

https://www.ust.com/en/insights/steering-at-speed-rebuilding-the-sdlc-for-an-ai-accelerated-world

https://www.ust.com/en/insights/ten-minute-take-from-jenkins-to-github-building-the-foundation-for-an-ai-native-sdlc

https://www.ust.com/content/dam/ust/documents/ust-and-github-partnership-solutions-final.pdf