Tech Radar: Techniques

Adopt - Context Engineering

Focuses on what information is available to agents through a dynamic pipeline (progressive context disclosure, prompt caching, dynamic retrieval, and context graphs) to improve reliability and scaling.

Adopt - Curated shared instructions for software teams

Shared operating manual front-loaded to the agent before each task, making agents more consistent and reducing quality variation.

Adopt - Dora Metrics

Metric for measuring software delivery performance (change lead time, deployment frequency, mean time to restore, change failure rate, and rework rate) to promote learning and productivity.

Adopt - Passkeys

Passwordless authentication using device origin cryptographic credentials, offering improved security and convenience.

Adopt - Structured output from LLMs

Forcing LLMs to output in specified formats like JSON to promote automation, reliability, and easier integration.

Adopt - Zero trust architecture

Security model that never trusts and always verifies every user and device using least-privilege principles and short-lived tokens.

Trial - Agent Skills

Capabilities or instructions applied to tasks to make agents more useful by loading only needed information. Caution advised for third-party skills.

Trial - Browser-Based Component Testing

Testing code directly in the actual execution environment, becoming increasingly consistent and useful.

Trial - Feedback sensors for coding agents

Mechanisms providing feedback from tests, reviews, errors, and runtime behavior to improve agent performance before code commit.

Trial - Mapping code smells to refactors techniques

Instructing agents to handle specific code issues with defined approaches or skills, though validation is still recommended.

Trial - Mutation Testing

Deliberately introducing bugs into source code to verify test suite effectiveness, acting as a safeguard for AI testing.

Trial - Progressive Context Disclosure

A context engineering technique giving agents a lightweight index of available data to pull only useful information, reducing token costs.

Trial - Sandboxed execution for Coding agent

Running AI-generated code in an isolated environment before affecting real systems to improve safety.

Trial - Semantic Layer

A shared business meaning layer defining concepts and metrics consistently across systems in an organization.

Trial - Server Driven UI

Moving UI decisions from client apps to server-controlled configurations to enable faster changes in dynamic application areas.

Assess - Agentic Reinforcement Learning Environments

Environments combining context, tools, and feedback to train agents on multi-step tasks, though reliable methods are still emerging.

Assess - Architecture Drift Reduction with LLMs

Using AI to detect structural and semantic violations and fix architectural drift in evolving codebases.

Assess - Code intelligence as agentic tooling

Giving agents direct access to tools aware of Abstract Syntax Trees (AST) to reduce token waste during refactoring.

Assess - Context Graph

Connected representation of relationships between code, docs, and systems to provide better AI context, though difficult to maintain.

Assess - Feedback Flywheel

Continuous retrospective-like improvement step for coding agents, though creating reliable automated feedback loops remains challenging.

Assess - HTML Tools

Portable and shareable scripts or utilities packaged as single HTML files, though they present security risks with untrusted files.

Assess - LLM evaluation using semantic entropy

Measuring uncertainty in LLM responses by analyzing answer variance to filter out false claims.

Assess - Measuring Collaboration Quality with Coding Agents

Evaluating how effectively humans and AI agents work together using specialized metrics beyond traditional developer metrics.

Assess - MITRE ATLAS

A knowledge base of threats and attack techniques targeting AI systems, important for security operations.

Assess - Ralph Loop

Feeding a fixed prompt to an agent in an infinite loop with fresh context to let it autonomously work against specifications.

Assess - Reverse Engineering for Design System

Using tools to extract design systems from legacy applications to aid modernization efforts.

Assess - Role-based contextual isolation in RAG

Architectural access control ensuring AI systems only retrieve context authorized for specific user roles.

Assess - Skills as executable onboarding documentation

Turning team knowledge into executable instructions for AI agents to preserve organizational knowledge.

Assess - Small language models

Compact AI models designed for specific tasks to help reduce operational costs.

Assess - Team of coding agents

Multiple specialized agents collaborating on software development tasks, pending better coordination techniques.

Assess - Temporal Fakes

Simulating real-world sequential systems for testing complex workflows.

Assess - Toxic Flow Analysis for AI

Identifying harmful or risky AI workflows and potential attack surfaces using tools like Agent Scan.

Assess - Vision language models for e2e document parsing

Multimodal AI models processing both images and text for document processing, pending higher reliability.

Caution - Agent Instruction Bloat

Accumulating excessive rules and prompts that breed unpredictability and maintenance overhead.

Caution - AI-accelerated Shadow IT

Building AI tools outside normal governance processes, creating security, compliance, and support risks.

Caution - Codebase cognitive debt

Gaps in developer understanding caused by AI generating the majority of the code, increasing long-term complexity.

Caution - Coding agent swarms

Deploying large numbers of agents on the same problem, resulting in heavy overhead and duplicated work.

Caution - Coding throughput as a measure of productivity

Misleadingly measuring developer output via lines of code or commit volume rather than actual value delivered.

Caution - Ignoring Durability in Agent workflows

Building agent systems lacking state recovery and persistence, leading to production unreliability.

Caution - MCP by default

Automatically integrating Model Context Protocol for every task, introducing unnecessary complexity.

Caution - Pixel-Streamed development environments

Remote development environments streamed as pixels, which can introduce latency and reliability issues.