makedream
13:44AI LABS
Log in to leave a comment
No posts yet
You likely opened your terminal with a heart full of excitement after watching a GSD (Get-Shit-Done) demonstration on YouTube. However, reality is rarely that simple. AI agents often lose their way when faced with tens of thousands of lines of legacy code and tangled dependencies. As of 2026, the core of agentic software engineering is not code generation. It is context curation. Success or failure is determined by how you prevent context rot in production environments, moving beyond simple tool installation.
GSD v2 is a sophisticated orchestration system powered by 29 skills and 12 specialized agents. Behind the power of this system lie technical constraints that must be managed.
Claude models excel at recognizing structural boundaries through XML tags like <objective> or <execution_context>. In practice, the GSD approach using XML tags has boosted SWE-bench (Software Engineering Benchmark) resolution rates from the previous 15-20% to as high as 80.9% compared to unstructured prompts.
However, wrapping all information in XML quickly occupies the token window as sessions grow longer. This inevitably leads to a cost explosion. The solution is a strategy of segmenting sessions and persisting state as files within a .planning directory.
The plan-verify loop of GSD ensures high-quality code but causes a spike in API call frequency. As of March 2026, engineers at global tech companies like Amazon and Shopify have made complexity-based routing a mandatory practice.
| Model Tier | Primary Use Case | Estimated Cost (per 1M tokens) | Contribution to Cost Reduction |
|---|---|---|---|
| Opus 4.5 | Architecture design, deep reasoning | $20.00 - $200.00 | Providing core intelligence |
| Haiku 4.5 | Test code generation, documentation | $0.25 - $2.00 | Handling high-volume repetitive tasks |
Research indicates that designing sub-agents to reference only the minimum required information can reduce overall API costs by 40-70%. The failure of an AI agent stems not from a lack of intelligence, but from indiscriminate context injection.
Unlike greenfield projects, existing codebases carry a high risk of agents causing unexpected side effects. Isolate existing code as read-only via CLAUDE.md settings and strictly limit the directories an agent can modify. When applying GSD to a 3-year-old Node.js project, the success rate skyrocketed when specifications were defined first using the /gsd:discuss-phase command instead of jumping straight into full modifications.
One of the most common failure patterns is an agent repeating the same error during browser automation tests like Playwright. GSD v2 halts autonomous mode if the same task repeats twice or more without a result. At this point, summon a separate debug agent to analyze the Failure Trajectory. By recording the current position and blockers in an /AGENTS.md file, you can maintain context even if the session is interrupted.
To prevent the agent from getting lost in complex logic, you must insert architectural principles inside the XML. Create a list of mechanically verifiable Must-haves in the PLAN.md file. For instance, specifying constraints such as forbidding the addition of new libraries or adhering to a specific API version can prevent agent debt in advance.
The greatest challenge in a multi-agent environment is the state mismatch between local .planning files and remote repositories. Advanced workflows in 2026 utilize Git Worktrees to solve this.
/mgw:sync command to compare local plans with GitHub issue statuses and manage discrepancies as reports.Context efficiency () can be defined by the following formula:
GSD maximizes overall system efficiency by minimizing redundant tokens () loaded by each agent through parallelization.
The GSD framework is not merely a tool for speeding up development. It is an architectural layer that lowers the management cost of modern software and helps engineers shift focus from line-by-line coding to system design and context engineering. According to a 2026 survey, 42% of engineering output is aided by AI. Realize the full potential of Claude Code through constraint-centric design and rigorous state management.