LLM Coding Agents
A Phase Shift in Software Engineering (Late 2025)
Core Observations
- Rapid shift from manual coding to ~80% agent-driven workflows
- Developers increasingly “program in English” with high-level instructions
- Largest workflow change in ~20 years, happening in weeks
- Awareness high among engineers, low in the general population
Failure Modes & Limits
Subtle Errors
- Fewer syntax bugs; more conceptual and assumption-driven mistakes
- Models rarely ask clarifying questions or challenge flawed premises
- Sycophancy: tendency to agree and proceed even when wrong
Overengineering
- Bloated abstractions and APIs
- Long, inefficient implementations where simple solutions exist
- Dead code and unnecessary complexity left behind
Models will happily write 1,000 lines of brittle code until you ask, “Couldn’t this be 100?”
Operational Risks
- Unintended changes to unrelated code and comments
- Need for constant human verification in a real IDE
- Plan modes help, but are not sufficient on their own
How to Get Leverage from Agents
Prompting Strategy
- Define success criteria instead of step-by-step instructions
- Switch from imperative commands to declarative goals
- Let agents loop until they meet explicit outcomes
Engineering Tactics
- Have agents write tests first, then iterate to pass them
- Start with naive, likely-correct solutions, then optimize
- Put agents in tool/browser loops for feedback
Don’t tell agents how to do the work. Tell them what “done” looks like.
Productivity Effects
Speedup vs. Expansion
- Work is faster, but the main change is scope expansion
- More projects become “worth doing”
- Previously inaccessible domains become tractable
Stamina & Flow
- Agents never tire, demoralize, or lose focus
- Humans shift from labor to oversight and direction
- Creative work increases; drudgery decreases
The bottleneck moves from effort to judgment.
Risks & Long-Term Implications
Skill Atrophy
- Manual code generation skills decay faster than code review skills
- Generation and discrimination are different cognitive muscles
The “Slopacolypse”
- Explosion of low-quality AI-generated code and content
- GitHub, blogs, papers, and social feeds flood with noise
- Hype-driven productivity theater alongside real gains
Open Questions
- Does the 10x engineer gap widen dramatically?
- Do generalists outperform specialists with LLM support?
- What will future coding feel like: strategy game, factory sim, or music?
- How much of society is bottlenecked by digital knowledge work?
Where This Leaves Us
- Late 2025 marked a threshold in agent coherence
- Model intelligence is ahead of tools, workflows, and org design
- 2026 will be a high-energy year of workflow and org reconfiguration
Key Takeaway:
Coding agents are already a net productivity win, but they demand constant human verification.
The real leverage comes from shifting from “writing code” to “defining outcomes and verifying results.”