Why the push for Agentic when models can barely follow a simple instruction?
a day ago
- #agent-management
- #python-coding
- #workflow-optimization
- Agents work best with structured folder systems and predefined .md files for context.
- Agents should ingest all relevant .md files before starting tasks to ensure they are well-prepared.
- Tasks should be broken into small vertical slices (under 100k tokens) to avoid agent misbehavior.
- Different specialized agents (architect, codeseeker, coder) can be used for specific tasks.
- Learning by observing agents and comparing their work is an effective way to improve coding skills.
- Agents must adhere to strict documentation and testing protocols (e.g., updating TASKS.md, WORKLOG.md).
- Code must be PEP8-compliant, typed, and thoroughly tested before submission.
- Agents should never execute code in terminals or REPLs; all changes must be file-based.
- Deterministic and complete solutions are required, with no partial work left behind.
- Agents follow a strict workflow: context gathering, task claiming, implementation, validation, and documentation.