Your agents lack context: Here's how to fix "You're absolutely right!" — Brandon Waselnuk, Unblocked

The Critical Role of Context in AI Code Generation and Agentic Systems

Brandon Waselnuk from Unblocked stresses that AI-generated code must feel as if it were written by a seasoned team member. This analogy underscores a core challenge: human developers acquire deep context through years of experience—attending meetings, shipping pull requests (PRs), and resolving production incidents—whereas AI agents begin each session with no institutional knowledge. This absence of rich, company-specific context presents a significant barrier to effective AI integration in software development.

Compounding Costs of Poor Context

The issue of inadequate context dramatically escalates as AI agents progress beyond basic tab-completion into more autonomous roles. Initially, with “tab-complete” models, humans act as an immediate context engine, quickly distinguishing good suggestions from bad. However, as agents are designed for increasingly independent operation, without direct human intervention, the cost of missing context compounds:

• Doom Loops: Agents become trapped in iterative correction cycles with human oversight, leading to wasted search tokens and significant rework.
• Review Tax: For parallel and background agents, insufficient context results in suboptimal code reviews, as the AI fails to grasp business logic or operational nuances.
• Ineffectiveness: Without proper context, agents cannot effectively query for information when encountering obstacles, hindering their ability to generate mergeable code, especially in complex brownfield projects.

Waselnuk highlights that just as early defect detection is crucial, addressing context issues early in the development lifecycle is essential to mitigate these compounding costs.

Common Pitfalls in Context Provisioning

Waselnuk identifies two prevalent “dead ends” in attempts to provide context to AI:

1. The Curated Context Trap:
   • Teams create markdown files or virtual file systems to document project context.
   • Challenges include: overhead in distribution, rapid documentation rot, and the unrealistic expectation of an “omnipotent” individual to maintain these resources for an entire organization.
2. The MCP Plateau:
   • This refers to agents utilizing external tools or Multi-modal Context Providers (MCPs).
   • Agent Failure to Call: Poorly written server or tool descriptions can lead to agents failing to invoke necessary external information sources.
   • Satisfaction of Search Bias: Agents often cease their search at the first plausible piece of information they find, overlooking more critical or recent context (e.g., a recent Slack discussion dictating a different approach from an older architectural record).
   • Ultimately, “access to information is not understanding.” An agent might generate code that compiles but causes production issues because it lacks crucial operational context, such as rollout procedures or feature flag management.

The Essential Functions of a Context Engine

To bridge the gap between mere information access and true understanding, a robust “context engine” is indispensable. Such an engine must:

• Understand Personal and Organizational Context: Comprehend the user’s identity, their role, and their operational environment (e.g., Git commits, code reviewers).
• Resolve Conflicts: Discern the authoritative source when information conflicts (e.g., old architecture diagrams versus recent CTO directives).
• Enforce Permissions and Governance: Integrate with security models (e.g., OAuth, SSO) to prevent unauthorized access to sensitive information.
• Deliver Token-Optimized Context: Provide relevant information efficiently, adapting output for both human (e.g., full detail) and machine (e.g., concise, token-optimized) consumption.

Inside the Engine: Six Key Characteristics

A context engine ingests diverse data from engineering sources (code, wikis, incident management) and real-time streams. It then processes this information based on six critical characteristics:

1. Unified System Context: Integrating data across all organizational systems to provide a holistic view for large enterprises. This enables models to identify “unknown unknowns.”
2. Targeted Retrieval: Balancing the demand for deep, comprehensive research with rapid, precise retrieval (e.g., quickly unfurling a link).
3. Conflict Resolution: Employing techniques to determine the correct information when conflicting sources exist.
4. Personalized Relevance: Tailoring context delivery based on the user’s identity, role, and current task.
5. Token Optimization: Ensuring that the context provided to the model is concise and efficient, thereby reducing token expenditure.
6. Permission Enforcement: Strictly adhering to access control policies, preventing unauthorized information leaks.

Demonstrable Impact: Context vs. No Context

Unblocked conducted a test, running the exact same prompt against the same model, once with a context engine and once without. The results were compelling:

• Without Context: Consumed 21 million tokens over approximately 4 hours.
• With Context: Consumed 10.8 million tokens and finished about 2 hours sooner.

This demonstrates a 50% reduction in token consumption and significant time savings, leading to faster triage and improved answer quality by “hydrating” the model with relevant, pre-processed context, thereby eliminating wasteful search tokens.

Open-Source Tools from Unblocked

Unblocked has released several open-source tools to assist teams in building and leveraging context:

1. Social Comment Network: A deterministic program that analyzes GitHub to map team members, their commits, and reviewers, generating an “experts graph.” It can optionally use LLMs to label teams, aiding in focusing context engines.
2. Repo Rules Agent: Discovers and indexes all rule files (e.g., linting, security policies) within a repository, identifies duplicates, and improves context retrieval for these crucial guidelines.
3. Workshop: “Building a Relational Context Engine from Scratch” (Beyond RAG): This workshop teaches how to extend beyond traditional Retrieval Augmented Generation (RAG) by enabling agents to discover schemas and write deterministic queries for relational data. This is crucial for answering complex questions like “What are the open PRs I worked on last week with authentication?” which RAG alone cannot handle.

Beyond Code Generation and The Future of AI

The applications of context engines extend far beyond code generation, benefiting areas like customer success (solving tickets faster) and sales (querying information in the field).

Waselnuk concludes that the primary gap in AI effectiveness is no longer intelligence but context. While models continue to improve, their true power in an organizational setting is unlocked by the quality and relevance of the context with which they are surrounded.

• Tags:
• AI Agents
• Code Generation
• Developer Tools
• RAG
• LLMs
• Context Engineering
• Enterprise AI
• Workflow Automation
• Token Optimization
• Unblocked