What could go wrong?

The hardest part of an ultra is knowing when to push, when to hold back, when the math stops mathing and instinct takes over. I learned this the hard way at mile 24 of a long training session, legs burning, clock ticking, watching my carefully constructed pace chart unravel with the climbs.

For my 50K this August, I had a plan. But as someone who builds and works with AI every day, I naturally wanted to explore how current models fail at meta-planning. So I brought in an agent. It failed spectacularly. The pacing chart bore no relation to reality.

The Problem

A 50K exceeds a long day in the forest. It is like putting together a puzzle upside down while blindfolded. Five aid stations, cutoff times, 8,500 feet of elevation change, sections of road and trail that demand different paces, and a 30-pound load on my back rewrite all the rules. Races are dynamic. Static plans fail. Human bias makes it worse. We overestimate early ability and underestimate late fatigue. Starting too quickly is a common pacing error in endurance events and can contribute to fatigue and reduced performance later in the race. The ones who finish are usually the ones who respected the constraints.

Enter the Agent

Recent work from ICLR 2025 on agent alignment drives home a point. Agents can plan only if we tell them. They lack built-in domain knowledge. They do not understand that a mile on trail is fundamentally different from a mile on road, or that sustained climbs burn legs in ways short bursts do not. They excel at structure, iteration, and synthesizing large amounts of data.

Here human and machine meet. The core issue was alignment. Agents lack the lived experience to translate human racing goals into effective segment-by-segment strategies. I provide the domain knowledge. The agent provides the hallucinations. Together, we struggled to build a workable plan.

The Case Study

Phase 1: The Naive Plan

I fed the agent the basics. Distance, aid station locations, cutoff times. The pacing chart was unrealistic. Each segment had a target time, though some of the math kind of worked.

Reality exposed the flaws. The chart predicted I would arrive at checkpoints after cutoff times, yet it claimed confidence. Other checkpoints left zero buffer. It assumed I could run all climbs. Road and trail were the same. Worst of all, it ignored the weight on my back.

Lesson: agents lack lived experience.

Phase 2: Surface Matters More Than Distance

I added surface data. The course has a clean breakdown.

• Road 0-0.5mi, 16.5-21.5mi, 29.5-31.7mi
• Trail 0.5-16.5mi, 21.5-29.5mi

The agent assumed trail miles take 20-26 minutes and road miles take 14-20 depending on grade. The agent recalculated. The plan improved marginally.

Lesson: the downhill road section from 16.5-21.5mi was fast and could be a solid time bank. Run it hard. Deposit minutes for later withdrawal. Duh.

Phase 3: Elevation vs. Slope Duration

I provided elevation profiles. The agent tallied total gain and loss but missed a critical subtlety: sustained grades break you, short steep bursts don’t. A 5% grade for three miles is harder than a 20% burst for 200 meters.

Segment 2 was the crux. 2,500 feet of gain over 3.1 miles. It was the most sustained. The agent called it hard. I called it survive.

Phase 4: The Buffer Epiphany

Next, it set 15-minute buffers for aid stations. It had no clue what actually happens at an aid station. Aid stations require check-in, refueling, gear changes, bathroom breaks, and mental resets. Fifteen minutes disappears.

We adjusted to 30+ minute buffers. The pacing slowed. The plan improved slightly from the first iteration. After burning through millions of tokens, I knew the wall of text the agent produced was not something my brain could parse mid-race.

Phase 5: The Cheat Sheet

Complexity kills execution. I asked the agent to distill the entire strategy into a one-page reference. It produced a table

Terrain	Pace (min/mile)	Strategy
Road downhill	14-16	Bank time
Road flat/uphill	18-20	Steady effort
Trail rolling	18-20	Maintain rhythm
Trail steep (>15%)	22-26	Power hike only

Five pillars:

1. Survive Segment 2
2. Run all downhill road
3. Power-hike steep trail
4. Bank time where possible
5. Maintain consistent fuel

The result was underwhelming. Besides thinking of how hard a 26 minute mile is on a 15% grade with 30 pounds on my back, I could have developed the plan the morning of the race. After wasting so much time and a metric ton of tokens, I came up with a better pace plan by hand.

What I learned about meta-planning

• Alignment Requires Domain Knowledge: This was outer alignment in action. I had to continuously inject domain knowledge because the agent’s training objectives did not reflect the realities of trail racing.

• Iteration Refines Alignment: Each iteration improved alignment between the agent’s output and my racing objectives. The first plan was completely misaligned. The fifth plan was mid at best.

• Simplicity Prevents Misalignment: Simple plans for humans are not simple for agents. The agent treated it as complex, creating more opportunities for misalignment. A simple one-page cheat sheet kept the lies visible.

• Constraints Enforce Alignment: Hard cutoffs acted as alignment guardrails, forcing decisions that served the overall race objective rather than local optimizations.

• Do It By Hand: Sometimes human-made is the best option to ensure alignment and not waste your time.

What This Means

For athletes: AI is still not better than coming up with a race strategy yourself. Sitting down to understand the race, the constraints, and how it will all come together is easier than letting agents produce unreliable slop.

For AI: A Survey on LLM-based Agents documents that agents with human guidance outperform solo LLMs by 2-3x on complex planning tasks. The gap between agent planning and human planning is domain knowledge injection. The day Frontier Labs stop offering $400k+ a year salary for program managers, maybe I will take a second look at agents for meta-planning.

The scenario illustrates the alignment problem in microcosm. The agent optimized for mathematical elegance while I needed survival. Without explicit alignment mechanisms (domain knowledge injection, iterative refinement, hard constraints), the gap between agent capability and human objectives remained unbridgeable.

The bigger picture reveals human-AI collaboration as alignment work. Each iteration narrowed the chasm between the agent’s bad ideas and my racing goals. The human provides context, intuition, and domain expertise. The agent stumbles through iterations, seeking compromise.

Just as in geology, where you read rock layers to understand the past, in ultras you read segments to predict the future. The agent could not resolve the uncertainty. However, working through the plan helped me structure my response.

Suffering is inevitable. A good plan makes it tolerable. A human plan makes it an experience.

---

See inline links for sources and further reading.