---
title: Planning and Reasoning — Can LLMs Actually Plan?
description: The AI planning debate rages on. We tested what LLMs can and cannot plan in itinerary construction — here is the honest answer.
canonical: https://nowah.xyz/blog/can-llms-actually-plan
lastModified: "2026-08-07T08:05:11.068Z"
---

# Planning and Reasoning — Can LLMs Actually Plan?

The AI planning debate rages on. We tested what LLMs can and cannot plan in itinerary construction — here is the honest answer.

This is the most honest piece I will write about AI capabilities. The AI planning debate has been running for years, with strong opinions on both sides. One camp argues that LLMs are fundamentally incapable of real planning because they are next-token predictors without world models. The other camp argues that scaling and chain-of-thought reasoning have effectively given LLMs planning capabilities that are "good enough" for most practical applications.

We have a unique vantage point on this debate because we ship an AI agent that plans real trips for real people. Not toy examples. Not benchmark tasks. Actual multi-destination itineraries with real constraints, real budgets, and real consequences for getting it wrong.

Here is the honest answer: LLMs can plan some things and cannot plan others. The boundary between the two is more nuanced than either camp admits. And the practical implications for AI trip planning are significant.

## The planning debate

![Illustration for this section](https://pics.nowah.xyz/website-media/ai-research-011-img-1.webp)

Yann LeCun, Meta's chief AI scientist, has repeatedly argued that LLMs cannot truly plan. His argument is that planning requires building a mental model of the world, predicting the consequences of actions, and evaluating multiple possible futures. LLMs do none of this. They predict the next token based on statistical patterns in their training data. Any appearance of planning is pattern matching, not reasoning.

The scaling hypothesis camp counters that the distinction between "real planning" and "very convincing pattern matching" does not matter if the outputs are functionally equivalent. If an LLM can produce a coherent travel itinerary that satisfies all constraints, does it matter whether it "truly planned" or just predicted what a good itinerary looks like?

My position, informed by building and evaluating a production travel agent, is that both sides are partially right and the nuance matters.

## What LLMs can plan today

Let me start with what works. LLMs are genuinely good at several planning-adjacent tasks in the travel domain.

**Simple itinerary construction.** "Create a 5-day itinerary for Paris" is a task that LLMs handle well. They can generate a reasonable day-by-day plan with appropriate activities, timing, and logistics. This works because the pattern is well-represented in training data. Millions of Paris itineraries exist on the internet. The model can synthesize a competent one.

**Constraint satisfaction for small constraint sets.** "Fly from SFO to NRT, departing April 3, returning April 10, budget under $1,500, prefer direct flights." The model can take these constraints, call the right search tools, and filter results appropriately. Six or seven constraints are manageable.

**Sequential booking coordination.** "Book a flight, then find a hotel near the airport for the first night, then find a different hotel in the city center for the remaining nights." The model can decompose this into steps and execute them in order, carrying forward relevant information (arrival time, airport location) between steps.

**Chain-of-thought reasoning about trade-offs.** "Should I book the direct flight for $650 or the one-stop for $420?" The model can reason through the trade-offs: time saved, comfort, fatigue, how the layover duration affects the overall travel day. It produces sensible analysis.

[Function calling](/blog/function-calling-breakthrough-enabled-agents) accuracy exceeds 95% for well-defined schemas, which means the mechanical execution of plans is reliable. The model selects the right tools and populates parameters correctly the vast majority of the time.

## Where planning breaks down

![Supporting diagram](https://pics.nowah.xyz/website-media/ai-research-011-img-2.webp)

Now for the failures. These are the cases where LLMs consistently struggle.

**Combinatorial explosion.** A 5-city trip (say, Rome, Florence, Venice, Milan, Amalfi Coast) with flexible dates, transport mode options between each city (train, bus, flight, rental car), variable day allocation, and a total budget constraint creates a search space with thousands of viable configurations. The model cannot evaluate all of them. It tends to satisfice: generate the first plausible configuration it can think of rather than exploring the space systematically.

**Temporal reasoning.** LLMs are surprisingly bad at temporal reasoning. "If I land at 11 PM, can I make a 7 AM train?" requires understanding overnight stays, transit from airport to hotel, morning preparation time, transit from hotel to station. The model often gets these calculations wrong or ignores them entirely.

**Resource allocation.** "I have 10 days and 5 cities. How should I allocate days?" requires weighing each city's attractions against travel time between cities against the user's interests. The model tends to default to equal allocation (2 days each) rather than reasoning about why you might want 3 days in Rome and 1 day in a smaller city.

**Long-range dependency management.** In a 20-turn [planning conversation](/blog/ai-trip-planning-one-conversation), the model might make a decision in turn 15 that contradicts a constraint from turn 3. The constraint is still in the context window, but the model's attention does not reliably track dependencies across many turns.

**Non-obvious interactions.** The model might not realize that booking a late-night flight to Venice means you need a hotel near the Venice airport, not in the city center, because water taxis do not run that late. This kind of domain-specific reasoning about real-world logistics is inconsistent.

## Decomposition as a workaround

The most effective strategy we have found for handling [complex planning](/blog/launching-multi-city-itineraries-complex-planning) tasks is decomposition: breaking a hard problem into easier sub-problems that the model can handle individually.

Instead of asking the model to plan a 5-city Italian trip as a single task, we break it down:

1. Determine the optimal city order based on geography and transport options
2. Allocate days to each city based on the user's stated interests
3. Search for inter-city transport for each segment
4. Search for flights for the first and last segments
5. Search for hotels in each city for the allocated dates
6. Assemble the pieces and verify that everything fits together

Each sub-task is within the model's planning capability. The assembly step catches inconsistencies. If the train from Florence to Venice arrives at 3 PM but the Venice hotel has a noon check-in, the agent flags the misalignment.

This decomposition approach works because it transforms a combinatorial planning problem into a series of constraint satisfaction problems, which is what LLMs are better at.

## Tools as planning amplifiers

Here is my strongest claim about LLM planning: tools compensate for most planning limitations.

The model does not need to reason about whether $450 is a good price for SFO-NRT in April. It queries the API and gets real data. It does not need to mentally model the Venice water taxi schedule. It queries a transport tool. It does not need to hold 50 flight options in working memory and compare them. The ranking tool does that.

Tools externalize the parts of planning that LLMs are worst at: numerical comparison, real-world fact checking, exhaustive option evaluation. The LLM handles the parts it is good at: understanding user intent, decomposing tasks, selecting which tools to call, and synthesizing results into coherent recommendations.

This is why I think the planning debate, while intellectually interesting, is somewhat academic for applied systems. The question is not "can LLMs plan?" The question is "can LLM + tools plan?" And the answer to that question is much more definitively yes.

## Empirical results

We track itinerary planning success rates by trip complexity.

Simple round-trip (1 city): above 90% success rate. The model handles this reliably.

Standard trip with hotel (1 city, flight + hotel): above 85%. Occasionally misses hotel-flight coordination details like check-in times.

Multi-segment (2-3 cities, same country): about 80%. Most failures relate to inter-city logistics or suboptimal time allocation.

Complex multi-city (4+ cities, cross-border): about 70%. Failures include missed [visa requirements](/blog/ai-agents-visa-requirements-documents), unrealistic transit times, and budget overflow from not tracking cumulative costs.

Open-ended planning ("plan my honeymoon, 2 weeks, $8,000"): about 60%. The model struggles with the unconstrained search space. It generates plausible suggestions but often misses creative options or makes assumptions that do not match the user's style.

These numbers improve steadily with each model generation. Chain-of-thought prompting helps. Decomposition strategies help. Better tools help. The trajectory is clear, even if we are not at human travel agent levels for complex trips.

Context windows of 200,000+ tokens ensure that the full constraint set is always available. The challenge is not memory. It is reasoning over that memory correctly.

## The roadmap

Several developments will improve LLM planning for travel.

**\[Reasoning models\]\(/blog/reasoning\-models\-complex\-travel\-decisions\)** that spend more compute on thinking before responding are specifically designed for multi-step planning tasks. Early results show meaningful improvements on constraint satisfaction problems.

**Search-augmented planning** combines LLM reasoning with systematic search over the option space. Instead of the model guessing at a good itinerary, it generates candidates and systematically evaluates them against constraints. This is closer to how humans plan complex trips: generate a few options, compare them, refine.

**Better tool design** gives the model better building blocks for planning. A "validate itinerary" tool that checks an entire plan for consistency is more useful than making the model catch inconsistencies through reasoning alone.

**User interaction as planning feedback** lets the model adjust its plan based on user reactions, effectively using the human as a planning oracle. "You suggested Rome for 2 days, but I really want to see the Vatican, the Colosseum, and Trastevere." The model learns that its allocation was too tight and adjusts.

## The user in the loop

There is one planning advantage AI agents have that pure AI planning benchmarks miss: the user is in the loop.

Academic planning benchmarks evaluate the model in isolation. Can the model, with no feedback, produce a correct plan? This is the hardest version of the problem.

In practice, the user provides feedback at every step. "Those dates work but can we fly in the morning?" "Florence gets more time, I want to see all the art." "That hotel is too far from the center."

Each piece of feedback constrains the search space and corrects planning errors. The model does not need to get the plan right on the first try. It needs to get it right iteratively, with human guidance.

This is a fundamentally different task than pure planning. It is collaborative planning, and LLMs are much better at it than the benchmark debates suggest. The user handles the creative judgment ("I want more time in Florence because the art matters to me"), and the model handles the logistical optimization ("Here is how we adjust the itinerary to give Florence 4 days while keeping the total trip at 10 days and the budget under $5,000").

## The compound planning system

Our itinerary planning is not a single model call. It is a compound system with multiple components.

**The model** handles intent understanding, conversational interaction, and high-level plan structure. It decides that a 3-city Italy trip should go Rome-Florence-Milan based on geography.

**The search tools** handle specific queries. Flight availability on specific dates. Hotel options in specific neighborhoods. Train schedules between cities. Each tool returns real data that the model incorporates.

**The constraint checker** validates the plan against all active constraints after each planning step. It catches inconsistencies: the Florence hotel checkout is at 11 AM but the train to Milan leaves at 9 AM.

**The cost tracker** maintains a running total of the trip cost, alerting the model when additions push the total beyond budget.

**The preference matcher** scores each plan element against the user's preference profile. If the model suggests a large chain hotel and the user prefers boutiques, the matcher flags the mismatch.

Together, these components produce better plans than any single one could alone. The model brings reasoning and creativity. The tools bring real data. The checkers bring consistency. The matcher brings personalization.

This compound approach is, I think, the practical answer to the planning debate. LLMs alone cannot plan complex trips reliably. LLMs as reasoning engines within a well-engineered system can. The system compensates for the model's weaknesses while leveraging its strengths.

## The reasoning model shift

The arrival of reasoning-focused models changes the planning picture meaningfully.

Standard language models generate text token by token, each token predicted based on the preceding context. This process is fast but does not naturally support the kind of look-ahead reasoning that planning requires. The model cannot easily ask itself "if I allocate 3 days to Rome, will I have enough time for Florence and Amalfi given the travel days between them?" without explicit prompting.

Reasoning models introduce an internal deliberation step. Before generating output, the model "thinks" through the problem, considering multiple options and their consequences. For itinerary planning, this means the model can evaluate alternative city sequences, compare time allocations, and identify scheduling conflicts before committing to a plan.

In our testing, reasoning models produce noticeably better first-draft itineraries for complex trips. The improvement is most visible in time allocation (more realistic day counts per city), inter-city logistics (better awareness of travel time between destinations), and budget distribution (more balanced allocation across trip segments).

The tradeoff is latency. Reasoning models take longer to generate output because of the deliberation step. For a simple flight search, this extra time is unnecessary. For a 3-city itinerary with budget constraints, the extra seconds produce a meaningfully better plan. We route to reasoning models selectively based on task complexity.

## Where pure planning still fails

Despite improvements, there are categories of planning where LLMs consistently struggle, and I want to be specific about them.

**Temporal arithmetic across time zones.** A trip from New York to Tokyo with a connection in London involves three time zones and a date line crossing. The model frequently makes errors calculating arrival times, layover durations, and local times at each stop. We offload this to a deterministic time zone calculator rather than trusting the model's arithmetic.

**Opportunity cost reasoning.** When the budget is tight, spending more on one component means less for another. The model can understand this conceptually but often fails to propagate the impact correctly. Adding a business class upgrade means the hotel budget needs to decrease, which means moving from a 4-star to a 3-star property, which changes the neighborhood options. The cascade of effects is hard for the model to track without tool support.

**Seasonal and event awareness.** The model might plan a trip to Barcelona in August without accounting for the extreme heat and the fact that many locals leave the city. It might schedule a Tokyo visit during Golden Week without warning about crowds and higher prices. This kind of contextual knowledge exists in the training data but is not reliably surfaced during planning. We supplement with a [seasonal intelligence](/blog/seasonal-intelligence-when-to-go) tool that flags known issues for specific destinations and dates.

These failure modes are not random. They are structural weaknesses in how language models reason. Recognizing them allows us to build targeted compensations: calculators for temporal math, constraint propagation tools for budget cascades, and knowledge bases for seasonal context.

## Testing planning quality

We test planning quality with increasingly complex scenarios.

**Scenario 1:** "Plan a week in Paris." Simple. One city, one set of dates. Success rate above 90%.

**Scenario 2:** "Plan 10 days in Italy: Rome, Florence, Amalfi Coast." Multi-city within one country. Need inter-city transport, time allocation, hotel coordination. Success rate about 80%.

**Scenario 3:** "Plan a 3-week trip through Japan during cherry blossom season for a family of 4 with a teenager and a toddler." Complex constraints: seasonal timing varies by region, family composition affects hotel and activity choices, long trip requires pacing and rest days. Success rate about 65%.

**Scenario 4:** "Plan my honeymoon. Two weeks. $10,000 budget. Surprise me." Open-ended. The model must generate destination ideas, evaluate them against the user's preference profile, and produce a complete itinerary. Success rate about 55%.

The success rate numbers are not embarrassing. A 65% success rate on a complex family trip to Japan means the agent gets it right two-thirds of the time on the first try. When it gets it wrong, user feedback in subsequent turns usually gets it to a good plan within 2-3 iterations. The effective success rate after iterative refinement is much higher.

## The trajectory

The planning debate is often framed as a binary: can LLMs plan or not? The more useful framing is directional. Are they getting better at planning, and how fast?

The answer is unambiguously yes. Each model generation shows measurable improvement on our planning benchmarks. Success rates on complex scenarios have improved from roughly 40% eighteen months ago to roughly 65% today. The improvement comes from better in-context reasoning, better tool use, and larger context windows that can hold more constraint information.

Extrapolating this trajectory is dangerous because progress is not linear. But even conservative projections suggest that LLM planning capabilities will be substantially better within a year. Combined with improved tool design and better evaluation frameworks, the effective planning capability of AI travel agents is on a steep upward curve.

We do not need to wait for LLMs to be perfect planners. We need them to be good enough planners that, augmented with the right tools and human feedback, they produce better outcomes than the alternative. The alternative is a user spending 38 hours across 45 browser sessions manually coordinating 6 bookings. That is a low bar, and AI trip planning already clears it for most scenarios.

Can LLMs actually plan? For travel, the honest answer is: they can plan simple to moderate trips well, they struggle with complex trips, and the combination of decomposition strategies plus external tools pushes the effective boundary considerably higher. The best AI trip planning today is not pure LLM reasoning. It is LLM reasoning amplified by good engineering. That is good enough to be useful. It is not yet good enough to replace a great human travel agent for complex trips. We are working on closing the gap, and based on the improvement trajectory we have measured over the past year, the gap is shrinking faster than most skeptics expect.

---

Nowah is an AI travel agent that searches and books real flights and hotels through conversation — no filters, no thirty open tabs. [Plan your next trip](https://app.nowah.xyz).