How Top Engineers Stop AI Agents From Writing Slop

How top engineers prevent AI agents from writing “slop” in codebases

High-level principle

Treat slop as an engineering problem, not an inherent LLM limitation. Modern LLMs can produce high-quality code if you build the right safeguards and workflows.

Primary rule

Never fix bad output in place. Diagnose, reset the run, fix the root cause, and rerun (or scrap the agent run) instead of patching agent-produced technical debt.

Core tools and how they’re used

• Hooks
  • Custom harnesses around agents for logging, pre-commit checks, stopping destructive actions, and traceability.
• Quality gates
  • Enforce strict linting and type-checking; require all tests to pass before work advances.
• Pre-commit / CI checks
  • Run tests, linters, and type checkers automatically when an agent attempts to commit.
• Hard blocks
  • Block risky commands/tools (for example, disallow git push from agents). Define read-only vs read-write capabilities (e.g., scout agents may only read).
• Per-agent isolation (worktrees)
  • Run each agent in an isolated working tree to prevent agents from overwriting or interfering with each other.
• Standardization
  • Standardize where issues/learning are recorded, prompt format, tool access, output naming, and review processes across agents.
• Traceability & logging
  • Record what agent changed what, when, and where so actions are auditable.

Testing & test philosophy

• Anti-mocking: avoid mocked tests because LLMs tend to produce shallow or misleading mocks. Favor integration/real tests that validate actual behavior.
• High coverage and a 100% pass rate are required before a change advances.
• If an agent writes failing tests, fix immediately or scrap the run and retry.

Task design and workflow techniques

• Task decomposition / focused agents
  • One agent — one task — one prompt. A focused agent is more reliable.
• Clear specs
  • Specs must remove ambiguity: exact file names, line numbers, code snippets, and expected behavior. More detail -> better agent output.
• Pit of success
  • Provide high-quality code and prompts so future agents produce higher-quality outputs (positive feedback loop).
• Multi-agent workflows / swarms
  • Chain agents for decomposition, implementation, and review. Require quality gates at each handoff so upstream agents never pass slop down the chain.
• Agent scope and scope enforcement
  • Explicitly define what files/areas an agent may touch and what’s outside its scope to reduce off-task behavior.
• Chain of command / human-in-the-loop
  • Standardize when developers are notified and where manual intervention is required.

Operational recommendations

• Standardize agent outputs, prompt structure, and tool usage across the codebase.
• Work as a team to define which tools/techniques to use — don’t leave this to ad hoc cloud code.
• If an agent produces bad output, prefer rerun after correction instead of incremental fixes that accumulate debt.

References / sources mentioned

• Stripe article (example of agents in production)
• Claude Code (example LLM/tool referenced)
• Indie Dev Dan: slogan “one agent, one task, one prompt”

Speaker / primary source

Jim West — Agentic engineer; builds publicly and shares resources (GitHub). Also points to his community Promptra and consulting offering for more engagement.