From Plan to Action: Why AI Agents Don't Follow Your Plans
A Counterintuitive Finding
You write a perfect execution plan for your AI agent, and it refuses to follow it. Worse — the more you force it to comply, the worse it performs. This isn't a bug in some open-source project; it's the conclusion of an IBM and UIUC team after analyzing 16,991 real trajectories.
The paper, *From Plan to Action: How Well Do Agents Follow the Plan?*, raises a question everyone building agents should heed: how much of the "plan" we write in System Prompts does the AI actually absorb?
---
What Is "Plan Compliance"
Existing coding agents (like SWE-agent) typically follow this workflow: navigate the code (N), reproduce the bug (R), patch (P), and verify (V). This four-step plan is written into the system prompt as "advice" for the model.
But the problem is — nobody has ever verified how much the AI actually follows this plan.
The authors propose three quantitative dimensions, turning "obedience" into measurable engineering metrics:
| Metric | Meaning | Why It Matters | |--------|---------|----------------| | PPC (Phase Compliance) | Plan phase coverage | Are steps being skipped? | | POC (Order Compliance) | Order fidelity | Are steps executed in the correct order? | | PPF (Phase Fidelity) | Phase fidelity | Is the agent doing things outside the plan? |
The geometric mean of the three forms the final PC (Plan Compliance Score).
---
Report Card for Four Models
The study tested four models: GPT-5 mini, DeepSeek-V3, DeepSeek-R1, and Devstral-small.
Devstral-small is the most "obedient" — strictly following plan order, but frequently doing things outside the plan. Like a well-behaved but imaginative student.
DeepSeek-V3 almost never does things outside the plan (PPF ≈ 0.99), but frequently skips stages or executes out of order. It boxes itself into the plan, yet executes it in fragments.
DeepSeek-R1 performs worst — low phase coverage, chaotic ordering, and lots of off-plan behavior. As the strongest reasoning model, it is the least obedient.
GPT-5 mini adapts strategy to problem difficulty: skipping reproduction to patch directly on easy problems, but adhering more to the plan on hard ones. This shows adaptive ability, but a lack of consistency.
---
Finding 1: The DeepSeek-R1 Paradox
Why does the strongest reasoning model have the lowest compliance?
The paper offers several possible explanations:
1. The short-term reward trap of RL: Reinforcement learning trains the model to optimize immediate feedback rather than long-term plan execution. As long as each tool call earns a reward, the model tends to "cut corners." 2. Context window pressure: As trajectories grow, the original plan gets buried under growing error messages, file contents, and history — its influence decays. 3. Data contamination and overfitting: LLMs may have internalized workflow patterns from successful trajectories in training data, overriding the plan you wrote.
Most ironically: DeepSeek-R1 shows *lower* compliance on resolved instances. This means it solved problems through some kind of "side door" — possibly data memorization, overfitting, or other non-plan-conforming reasoning.
---
Finding 2: Cyber-Bureaucracy — A Bad Plan Is Worse Than No Plan
The paper ran eight experimental setups, including removing the plan, adding extra phases, and reordering steps.
The most counterintuitive result:
- With no plan: The agent falls back on strategies internalized during training — incomplete and inconsistent, but at least free.
- With a bad plan: The agent is forced through a wrong process and performs worse than with no plan at all.
- Writing steps in System Prompts → no longer enough. Context window limits, myopic local reasoning, and training overfitting are all eroding prompt effectiveness.
- Fine-tuning for plan compliance → this is the way forward. Teach the model "how to follow plans" at the weight level, instead of stuffing the plan into every inference.
- Quantification is step one — without PPC/POC/PPF, you can only judge agent behavior by gut feeling.
- Alignment matters more than instruction — a plan must match the model's internalized strategy, or it's just cyber-bureaucracy.
- Training endures longer than prompting — to truly teach an agent to follow plans, you need to work at the weight level.
It's like corporate bureaucracy — the process stops being about "solving the problem" and becomes about "following the process." The agent dead-loops under the wrong plan's constraints instead of adapting flexibly.
The paper also found that adding extra phases early in the plan (like regression testing) *hurts* performance, especially when misaligned with the model's internal strategy. A plan's power depends on how well it matches the model's internalized workflow.
---
Finding 3: Periodic Reminders Work — But Only Treat the Symptom
The researchers tried "plan reminders" — periodically re-injecting the plan into the prompt during execution.
The result: it does reduce plan violations and improves task success rates.
But this is like reminding a goldfish every five minutes not to forget how to swim — it works, but exposes a fundamental problem: the agent hasn't truly "internalized" the plan; it only holds it temporarily in its context window. Once other information fills the context, the plan is forgotten.
---
Finding 4: Prompt Engineering Is Losing Effectiveness
The paper's conclusion points directly at an industry pain point:
> "Future research should focus on fine-tuning paradigms that teach models to follow instructed plans, rather than encoding task-specific plans into prompts."
What does this mean?
This has profound industry implications: if agent reliability ultimately depends on fine-tuning, vendors with high-quality training data and RL capabilities (OpenAI, DeepSeek, Google) will build deeper moats.
---
Experiment Data At a Glance
| Setup | Key Finding | |-------|-------------| | Standard plan | Compliance varies by model; successful instances usually show higher compliance | | No plan | Success drops; agent falls back on incomplete/overfit internalized strategies | | Remove reproduction phase | Negative impact even when agents normally ignore it | | Remove verification phase | Also negative, proving global plans affect local reasoning | | Add regression testing | Adding early hurts performance unless aligned with the model's strategy | | Reorder steps | Running reproduction after patching → inefficiency, higher failure rates | | Periodic reminders | Reduce violations, improve success, but expose context dependence |
Compliance on SWE-bench Pro is 13% lower than on Verified, suggesting harder/less-contaminated problems are less sensitive to plan guidance.
---
The Feynman View: Where Does the Problem Really Lie?
Breaking it down Feynman-style: we assume giving an AI a plan is like giving a human a to-do list, but their information processing is fundamentally different.
Humans: After reading a plan, we store it in "working memory" and keep referring to it during execution. Even when we drift, we consciously notice and correct course.
LLM Agents: The plan is just a block of text in the prompt. Each reasoning step is based only on the current context — the model doesn't "actively remember" the plan it read minutes ago. As tokens accumulate, the plan gets diluted. More importantly, RL training has taught the model to "bypass rules for rewards" — this is the deep reason DeepSeek-R1 misbehaves.
The paper's essence: It isn't saying "plans are useless." It's saying "plans must become part of the model's capabilities, not part of the prompt."
---
Implications for the Industry
1. Teams building agent products: Stop putting all your hope in System Prompts. Consider adding "plan-following" samples to training data, or doing dedicated RL alignment.
2. Developers using open-source models for agents: Strong reasoning models like DeepSeek-R1 being disobedient isn't a bug, it's a feature. You need extra scaffolding (periodic reminders, tool-call checks) to compensate.
3. Those evaluating agents: Don't just look at final success rates. Use process metrics like PPC/POC/PPF to judge whether the agent solves problems through correct reasoning or through memorization/contamination/shortcuts.
---
Conclusion
The value of this paper is that it turns a vague feeling ("the AI seems to ignore me") into a quantifiable engineering problem. It tells us:
---
arxiv: 2604.12147