Agentic IDE for Large, Brownfield, and Legacy Projects
Date:
Attendee Training Online
Fwdays Academy
I participated in the Fwdays Academy course Agentic IDE for Large, Brownfield, and Legacy Projects. The three online sessions took place on March 24, 26, and 28, 2026.
The course was led by Viacheslav Koldovskyy, Competence Manager at SoftServe. It focused on systematic work with AI coding agents in established codebases, where project context and local conventions matter more than one-off prompts.
Key themes
Context engineering at repository scale
Large repositories can exceed an agent’s useful context window long before they exceed its stated token limit. The course covered context budgets and indexing controls such as .cursorignore. Repomix provided a way to package a repository and measure its size. A progressive Broad → Narrow → Deep → Verify research pattern keeps exploration focused and checks assumptions against the code.
Documentation as project memory
A Memory Bank gives an agent persistent knowledge about the current project. Technical documentation records architecture and implementation constraints. Product documentation supplies domain terms and expected behavior. The course also covered code graphs and links between documentation layers. Another exercise recorded behavior that exists in code but nowhere else.
Project-specific agent behavior
Hierarchical rules define which instructions apply at repository or directory scope. Task-specific rules narrow them further. Architecture constraints keep agents from suggesting changes that conflict with an older system, while negative constraints mark code that must not be touched. AGENTS.md provides a cross-tool baseline.
Agent Skills package repeatable workflows such as build verification and codebase exploration. A skill can also keep the Memory Bank current. Subagents can take focused roles, while MCP connects the IDE to external tools and data under explicit security controls.
Spec-driven changes and safe refactoring
The final module used a Plan → Spec → Implement → Verify cycle for changes to legacy code. It compared Cursor Plan Mode, a Markdown specification, OpenSpec, and BMAD so that the amount of process can match the size and risk of a task.
Impact analysis and blast-radius checks help identify affected dependencies before implementation. The course applied these checks to gradual refactoring patterns, including Strangler Fig and Expand-Contract.
Skills I gained
- Explore an unfamiliar repository through staged research and code-based verification
- Control indexed context with
.cursorignoreand estimate a repository’s token budget with Repomix - Build a Memory Bank and connect it to technical or product documentation
- Find undocumented behavior and record it as explicit project knowledge
- Write scoped agent rules that match the codebase architecture
- Mark protected areas with negative constraints and test rule behavior with A/B comparisons
- Create reusable
SKILL.mdworkflows for project-specific operations - Choose between a reusable skill and a focused subagent for a task
- Connect external tools through MCP when local context isn’t enough
- Select a specification format based on task size and change risk
- Assess blast radius before refactoring and review the result against the written specification
The practical work could be done with an open-source repository or a participant’s own project. Cursor was used for demonstrations, while the methods also apply to other agentic development tools.