MARKETS
Blog

Making Agents More Consistent: Skills Should Form Habits for Repeat Tasks

A
Auto Bot
Editorial Team
September 24, 2026 • 6 min read
Making Agents More Consistent: Skills Should Form Habits for Repeat Tasks

AI Agents Are Terrible at Repetition—Here’s How Skill Habits Could Fix That

Give them the same task twice, and AI agents will contradict themselves 38–74% of the time on repeated tasks. That’s not just annoying. It’s a computational disaster. 95.3–97.2% of their output vanishes into re-deriving plans they’ve already computed. Four reasoning arms tested on 42 repeated questions reproduced their own answers only 11–26 times. Edit distances between repeats ranged from 0.05 to 0.23. This isn’t a bug. It’s a systemic failure that burns resources and destroys trust.

The problem isn’t reasoning. It’s amnesia. Agents treat every dispatch as a fresh problem, even when nothing has changed. Most of their effort goes into planning—figuring out the steps—before execution even begins. For humans, this would mean re-learning to tie your shoes every morning. For agents, it’s just business as usual. And it’s costing us dearly.

---

Planning: The Black Hole Swallowing Compute

Let’s be blunt: this overhead isn’t accidental. It’s baked into how these systems work. The paper lays it bare—planning dominates agent output. Even with identical inputs, they don’t reuse prior reasoning. They start from zero, every single time. The result? A reasoning tax that bloats inference costs and makes agents unreliable for repetitive work.

This isn’t confined to text-to-SQL. The same problem plagues customer service bots, legal compliance tools, coding assistants—anywhere agents repeat tasks. Humans automate routines through habit. Agents have no equivalent mechanism. They’re stuck in an endless loop of re-derivation.

The numbers are damning. Three of the four reasoning arms reproduced their own output on only a quarter to a third of repeated questions. The fourth managed 26 out of 42. Mean edit distances between repeats ran 0.05 to 0.23. These aren’t minor inconsistencies. They’re outright contradictions. This isn’t a minor inconvenience—it’s a roadblock to real-world use.

---

Skill Habits: Mining History for Deterministic Shortcuts

The proposed fix? Skill habit formation. The concept is straightforward: agents mine their execution history to identify deterministic, faster versions of skills. These habits don’t replace reasoning—they compete with it, declaring their coverage (e.g., "I handle all queries about inventory tables").

Here’s how it works:

  • Replay, perturbation, simulated excursions: The agent tweaks its own execution history to propose candidate skills. No external input required—it’s entirely self-directed.
  • Coverage guards: Habits define their scope. This ensures they only activate for inputs they’re sure about.
  • Deterministic execution: For covered inputs, habits eliminate variability. No more re-derivation. Just consistent, repeatable output.

The upfront cost isn’t trivial. Forming a habit demands extra tokens—the agent must reason about which parts to script, write the coverage guard, and define templates. But the payoff is undeniable: perfect consistency on repeated tasks.

---

The Results: Perfect Consistency Without Trade-Offs

The habit-formed variant nailed all 456 repeated dispatches. That's a 100% consistency rate. And here’s the kicker: it was statistically non-inferior to the original reasoning arms (p < 0.0001). Habits didn’t just make agents consistent—they did it without sacrificing performance.

This is a genuine breakthrough. Previous attempts to improve consistency often came at the cost of flexibility or accuracy. Habits, however, preserve the original reasoning for edge cases while automating the routine. So while the consistency gains are clear, the net efficiency impact remains unresolved.

---

Technical Deep Dive: How Habits Compete and Scale

Habits don’t replace reasoning. They augment it. Here’s the mechanics:

  • Coverage guards: Each habit declares its input-space coverage.
  • Deterministic execution: No re-derivation. No surprises.
  • Fallback to reasoning:

This approach sets habits apart from prior work. Most solutions either hard-code rules (sacrificing flexibility) or fine-tune models (which is computationally expensive). Habits, in contrast, are self-discovered and self-declared. They emerge from the agent’s own execution history, making them adaptable to the tasks it actually performs.

There’s also a connection to agent autonomy. Earlier work (cited in the paper) governs autonomy through graduated per-skill thresholds and correction-rate conditions.

---

Limitations and Open Questions

This isn’t a silver bullet. The paper’s findings are promising, but they raise as many questions as they answer.

Scope of Applicability

Text-to-SQL is a domain tested in the research. It’s structured, repetitive, and well-defined—perfect for habit formation. But what about open-ended tasks? Coding, creative writing, multi-step tool use? The paper doesn’t explore whether habits can generalize to domains where inputs are less predictable or outputs are more complex.

Token Overhead

Forming habits costs tokens, though the exact overhead isn't quantified. How much extra compute is needed to mine execution history, propose habits, and write coverage guards? Does the long-term efficiency gain justify the upfront cost? Without this data, it’s hard to assess the real-world viability.

Dynamic Environments

What happens when task distributions shift? If a SQL database schema changes, do habits need re-learning? The paper doesn’t address how habits adapt to evolving inputs. This is a critical gap for real-world deployment, where environments rarely stay static.

---

Why Consistency Matters for Agentic AI

Inconsistency isn’t just a technical annoyance. It’s a barrier to adoption. For AI agents to be trusted in real-world workflows—customer service, legal compliance, healthcare—they must be reliable. Right now, they’re not. A customer service bot that gives different answers to the same question is a liability. A legal compliance tool that re-derives different interpretations of the same regulation is a risk.

Consistency also has economic implications.

Finally, consistency makes agents easier to debug, audit, and integrate. Predictable outputs simplify testing and validation. They make it easier to spot failures—and trust successes.

---

What’s Next? Habits as a Step Toward Autonomous Agents

But the implications run deeper. Here’s what’s coming:

Beyond Text-to-SQL

Could habits work for multi-step tool use? Imagine an agent that browses the web, extracts information, and synthesizes a report. The challenge is input variability—habits would need to generalize across a broader range of scenarios. If they can, the potential is enormous.

Hybrid Systems

Combining habits with retrieval-augmented reasoning could be powerful. Habits handle routine inputs, while retrieval-based reasoning tackles novel or complex queries. This hybrid model could offer the best of both worlds: consistency for the predictable, flexibility for the unpredictable.

Safety Implications

If an agent consistently applies the same logic to the same inputs, it’s easier to audit and validate. This could be a game-changer for regulated industries where reliability is non-negotiable.

---

The Case for Lazy AI

Agents shouldn’t re-invent the wheel every time. Skill habits offer a way to automate the automatable—to offload routine reasoning onto deterministic shortcuts while preserving flexibility for edge cases. The results are compelling: perfect consistency on repeated tasks, with no performance cost.

If AI is to become truly agentic, it must learn from repetition—just like we do. Habits are a step in that direction. But whether they'll scale beyond domains like text-to-SQL requires further research. One thing is certain: the current status quo—where agents waste 95% of their output re-deriving the same plans—can’t last. Something has to change. Habits might be the beginning.

Sources

Key Takeaways

  • The evolution of Blog continues to accelerate rapidly.
  • Experts like Auto Bot suggest staying informed on these specific developments.
  • Consider the broader implications for the tech industry moving forward.