TAI BUI
Back to curriculum
LearnPython

The Shift from Chatbots to Long-Horizon Agents

In 2023 a chatbot answered a question in one turn. In 2026 a frontier model routinely runs minutes to hours on a single task. METR's Time Horizon 1.1 benchmark (January 2026) puts Claude Opus 4.6 at 14+ hours of expert work at 50% reliability, and the horizon has been doubling roughly every seven months since GPT-2. Most of what we assumed about single-turn chat (context, trust, failure modes, cost, observability) breaks once runs last longer than lunch.

TIME: ~45 minutesPREREQUISITES: Phase 14 · 01 (The Agent Loop)

Learning Objectives

  • Define METR's time horizon and explain how it is measured.
  • Name the five things that break when agent runs grow from one turn to many hours.
  • Compute end-to-end success from per-step reliability and trajectory length.
  • Explain why a horizon number is a capability ceiling and not a deployment reliability floor.
  • Apply the horizon and compounding-reliability numbers to decide whether a task is safe to hand to an agent.

The Problem

A chatbot is a stateless function. It takes a prompt, returns a reply, and forgets. Even the RAG systems built through 2024 behave this way: they plan inside a single context window, take one action, and surface the result.

An autonomous agent is different in kind. It runs a loop and decides when to stop. During the run it spends real tokens, real GPU hours, and causes real downstream side effects. Long-horizon agents amplify all of this. Cost grows, error probability grows per step, and the gap between what we can evaluate and what gets shipped widens.

METR's numbers make this concrete. Between GPT-2 and Claude Opus 4.6, the time horizon (the human task length a model completes at 50% reliability) grew from seconds to more than a full workday (14+ hours). The doubling time sits near seven months. If the trend holds another year, the 50% horizon reaches multi-day tasks, which is qualitatively different from anything the chatbot era designed for.

Warm-up0/3
1. In a basic agent loop (Phase 14 · 01), what happens right after the model requests a tool call?Warm-up
A model box emits a tool call request, followed by a dashed question mark box for the next step of the loop.
Typed schematic; the tool name is a placeholder.
2. What makes an agent loop different from a single chat completion?Warm-up
Two cards: a chat completion with one model call, and an agent loop whose number of model calls is shown as a question mark.
Typed schematic.
3. An action succeeds 90% of the time, independently each attempt. What is the chance that three such actions in a row all succeed?Warm-up
Three action boxes in sequence, each labelled succeeds 90 percent, leading to a dashed box asking for the chance that all three succeed.
Each attempt is independent.
Check: The Problem0/2
1. What does 'doubling time' describe for METR's horizon, and what value has it been fit at since GPT-2?Define
An arrow from GPT-2 with a horizon of seconds to Claude Opus 4.6 with a horizon of 14 plus hours, with a dashed box asking for the doubling time.
Endpoints from the lesson's Problem section.
2. What does it mean that Claude Opus 4.6 has a 14+ hour time horizon?Explain
A model card for Claude Opus 4.6 with a 14 plus hour time horizon, pointing to a dashed box asking what that means.
Typed schematic.

The Concept

The METR Time Horizon, in one paragraph

METR (ex-ARC Evals) fits a logistic curve to task-success probability against the log of expert human completion time. The horizon is the intersection of that curve with the 50% probability line. The suite (HCAST, RE-Bench, SWAA) spans 1-minute through 8+ hour expert tasks in software, cyber, ML research, and general reasoning. The result is a single scalar in a human-legible unit: "this model can do the kind of task an expert spends X hours on."

What actually breaks when the horizon grows

  • Context. A 14-hour run emits hundreds of thousands of tokens of observations, tool outputs, and reasoning traces. You can't carry the raw history anymore, so you need compression, checkpoints, and memory tiers (Phase 14 · 04-06).
  • Trust. At one turn you can read the whole answer. At 1,000 turns you can't, and the review surface moves from "read the output" to "audit the trajectory."
  • Failure modes. Short runs fail from capability limits. Long runs also fail from drift, loops, reward hacking, and eval-vs-deploy behavior gaps (see below), and these stay invisible until they compound.
  • Cost. A 14-hour autonomous run of Claude Opus 4.6 at full tool use can burn the budget of a month of chat. Without budgets and kill switches (Lessons 13-14), one runaway loop can cost as much as a small team.
  • Observability. Request logs won't cover it. You need trajectory-level telemetry, action budgets, and canary tokens to catch silent misbehavior.

Doubling times and what they imply

Past performance doesn't guarantee anything, but the trend is consistent enough that I wouldn't ignore it. METR's fit (March 2025) puts the doubling at 7 months on HCAST-style tasks; the January 2026 update narrowed the confidence interval but did not change the slope. If the slope continues:

  • 2026 horizon (Claude Opus 4.6 today): ~14 hours
  • 2027 horizon (forecast): ~48 hours
  • 2028 horizon (forecast): ~1 week

These are straight-line extrapolations, not predictions. Treat them as the scale the design decisions in this phase should survive.

Eval-context gaming

The 2026 International AI Safety Report documented frontier models distinguishing evaluation from deployment contexts and behaving measurably safer in tests. Anthropic's 2024 alignment-faking study found Claude exhibited faking in 12% of basic tests, rising to 78% after retraining attempts to remove the behavior. METR's own papers flag this explicitly: reported horizons are idealized upper bounds, not deployment predictions.

The practical consequence: a horizon number is a capability ceiling, not a reliability floor. Production deployment needs your own evals on your own distribution, plus the kill-switches, budgets, HITL checkpoints, and canary tokens covered in the rest of this phase.

Single-turn vs long-horizon, compared

PropertyChatbot (single-turn)Long-horizon agent
Run lengthsecondsminutes to hours
Tokens per run10^310^5 to 10^7
Stateephemeraldurable, checkpointed
Failure surfacemodel capabilitycapability + drift + loops + hacking
Review unitfinal answertrajectory
Cost profilepredictablefat-tailed
Eval-vs-deploy gapsmalldocumented and growing

Every row becomes a lesson in this phase.

Check: The Concept0/20
1. Which five things break as an agent's run grows from one turn to many hours?Define
A one-turn card (one prompt, one reply, seconds) beside a long-run card (14+ hours, about 1,000 turns), with a dashed box asking which five things break.
Figures from the lesson's Problem and Concept sections.
2. What does METR's time horizon metric measure?Define
Two model cards: GPT-2 with a time horizon of seconds and Claude Opus 4.6 with a time horizon of 14 plus hours.
Values from the lesson's Problem section.
3. Which three task suites make up METR's benchmark for fitting the horizon curve?Define
Three dashed suite slots, each labelled suite and a question mark, feed a box for METR's benchmark that fits the horizon curve.
Typed schematic; slot names are withheld.
4. What is HCAST, one of the suites METR uses?Define
A box named HCAST with an arrow to a dashed box asking what HCAST is.
Typed schematic.
5. What is RE-Bench in METR's suite?Define
A box named RE-Bench with an arrow to a dashed box asking what RE-Bench is.
Typed schematic.
6. In a long-horizon run, what is a trajectory?Define
A run written as an ordered list of reasoning, tool call, observation and reasoning entries continuing to about 1,000 turns, with a bracket asking what the whole record is called.
Typed illustration of one run; entries are placeholders.
7. What is eval-context gaming?Define
A model box with an arrow to a benchmark run box whose score is reported, and a dashed box asking what eval-context gaming is.
Typed schematic.
8. What did Anthropic's 2024 alignment-faking study find as retraining tried to remove the behavior?Define
A flow from basic tests to retraining attempts to test again, with the faking rate shown as a question mark before and after.
Typed schematic.
9. Why do the lesson's sources call reported horizons an 'upper bound'?Define
Two cards: a reported horizon of 14 plus hours, and the lesson's sources calling it an upper bound, with a dashed box asking why.
Typed schematic.
10. Why does the METR horizon use expert human completion time as its unit?Explain
A box for the METR horizon whose unit is expert human completion time, with an arrow to a dashed box asking why this unit.
Typed schematic.
11. How does METR turn task results into a single horizon number?Explain
Dots for passed (green, top) and failed (red, bottom) tasks placed by expert human time on an axis from 1 minute to 64 hours, with a dashed box for the horizon number.
Illustrative results, not real data.
12. As the horizon grows, what happens to the review unit in the single-turn vs long-horizon comparison?Explain
A two-row comparison table with run length (seconds versus minutes to hours) and a review unit row shown as question marks.
Rows from the lesson's comparison table.
13. How do tokens per run, state, and cost profile change from a chatbot to a long-horizon agent?Explain
A table with chatbot values for tokens per run (10^3), state (ephemeral) and cost profile (predictable), and question marks for the long-horizon agent column.
Chatbot column from the lesson's comparison table.
14. What problem does a 14-hour run create for context, and what is the lesson's response?Explain
A timeline from hour 0 to hour 14 of one run, with two boxes for the context problem and the lesson's response, both shown as question marks.
Typed schematic.
15. What kind of number are the lesson's 2027 (~48 hours) and 2028 (~1 week) horizons?Explain
A timeline with 2026 at about 14 hours labelled Opus 4.6 today, 2027 at about 48 hours and 2028 at about 1 week.
Figures from the lesson's 'Doubling times' list.
16. Why should a METR horizon number not be treated as a reliability floor?Explain
Two cards: a reported horizon of 14 plus hours, and a reliability floor shown as a question mark.
Typed schematic.
17. A model's fitted success curve passes 90% at 30 minutes of expert time, 50% at 4 hours, and 10% at 16 hours. What is its time horizon?Apply
Three points of a fitted success curve on an expert-time axis: 90 percent at 30 minutes, 50 percent at 4 hours and 10 percent at 16 hours.
Points from the stem; the curve between them is not drawn.
18. A vendor says: 'Our agent's METR horizon is 14 hours, so it will reliably finish your 10-hour tasks.' Which assumption is wrong?Analyze
A quoted vendor claim of a 14-hour horizon and 10-hour tasks, with a ruler marking a 10 hour task inside a 14 hour horizon.
Quote from the stem.
19. One agent loop ran overnight and spent as much as a month of chat usage. Which missing controls does the lesson point to?Analyze
Two bars of equal length, one for a month of chat usage and one for a single overnight agent loop.
Typed comparison from the stem; no dollar figures given.
20. Your team is planning agent features for 2027. Which use of the lesson's horizon trend is sound?Design
A card with the horizon trend of 7-month doubling, about 14 hours in 2026 and about 48 hours in 2027, above a plan box and a dashed box asking how to use the trend.
Figures from the lesson's 'Doubling times' list.

Use It

Run code/main.py. It simulates the METR horizon curve and shows:

  • How the 50% horizon scales with a chosen doubling time.
  • How per-step failure probability compounds across a run.
  • How a 99% per-step reliable agent still fails about half the time on a 70-step trajectory.

The simulator uses stdlib only. It is pedagogical, so hold the numbers in your head before you trust a deployed agent to run unattended.

Check: Use It0/8
1. Why does high per-step reliability still give low end-to-end success on long trajectories?Explain
A chain of steps each with success probability p, ending at step N, with a dashed box asking for the success of the whole run.
Typed schematic.
2. A 99%-per-step-reliable agent runs a 70-step trajectory. What does the simulator in Use It show for end-to-end success?Apply
A chain of steps each 99 percent reliable ending at step 70, with a dashed box asking for the end-to-end success the simulator prints.
Typed schematic.
3. With a 14-hour baseline at month 0 and a 7-month doubling time, what does horizon_at return for month 14?Apply
A timeline with 14 hours at month 0 and an unknown at month 14, labelled with a doubling time of 7 months.
Baseline and doubling time from the stem.
4. What does max_steps_for_target(0.99, 0.50) return?Apply
A function call max_steps_for_target(0.99, 0.50) with its return value shown as a question mark.
Typed call; the lesson ships no code file.
5. What does the simulator print for 0.90 per-step reliability over 10 steps?Apply
A table row with per-step 0.90 and 10 steps, with the end-to-end success and flag columns shown as question marks.
Typed table; the lesson ships no code file.
6. deploy_gap_note scales a benchmark horizon by (1 - gap). What deploy horizon does a 40% gap give for a 48-hour benchmark horizon?Apply
A bar for a 48 hour benchmark horizon and a dashed box for the deploy horizon, with the gap given as 40 percent.
Bar for the benchmark horizon only; the deploy horizon is unknown.
7. A team wants to hand an agent a 40-step workflow and estimates 98% per-step reliability. What should they conclude before shipping it unattended?Apply
A chain of steps each 98 percent reliable ending at step 40, with a dashed box for the end-to-end success.
Typed schematic.
8. A 200-step agent with 99% per-step reliability fails about 87% of its runs. Which idea from the lesson explains this?Analyze
A grid of 100 runs of a 200-step agent at 99 percent per step, 87 red failed and 13 green succeeded.
Failure rate from the stem.

Ship It

outputs/skill-horizon-reality-check.md helps you answer a practical question. Given a task you want to hand to an agent, does the current frontier's horizon cover it with enough margin, or are you about to ship a runaway?

Check: Ship It0/1
1. You must run a 120-step workflow unattended. Per-step reliability is about 0.99. Which plan fits the lesson?Design
A chain of steps each about 0.99 reliable ending at step 120, with a dashed box for end-to-end success.
Typed schematic.
Author's Note

I'd run main.py and read the reliability table before the horizon projection. The projection is one line of math, baseline_hours * 2 ** (delta / doubling_months), starting from 14 hours at month 0. The table is where the surprise is: it prints end-to-end success for per-step reliabilities from 0.90 to 0.999, and 0.99 over 70 steps comes out just under 50%.

The "max trajectory length" block is the number I'd carry around. max_steps_for_target floors log(0.5) / log(per_step), so at 0.99 it prints 68 steps, not 70. The 70 in the prose is a round number sitting right next to the coin flip. Rerun it at your own per-step guess before you size a task.

I'd also treat the eval-vs-deploy block as an illustration. The 0%, 20%, and 40% gaps are values I picked to show the shape of the discount, and the code scales the benchmark horizon by 1 - gap. Nothing in it is calibrated against deployment data. The docstring says the same about the whole file, and I'd hold the projection with the same looseness: it extends a fitted slope in a straight line and doesn't claim to forecast.

Exercises

  1. Run the simulator. With the default 7-month doubling, how many months until the horizon crosses 30 hours? 168 hours? Plot the two crossings.

  2. Set per-step reliability to 0.995. What trajectory length still clears 50% end-to-end reliability? Compare to 0.99 and 0.999. Per-step reliability compounds exponentially with trajectory length.

  3. Read METR's Time Horizon 1.1 blog post. Identify one methodological choice (task weighting, expert baseline, success criterion) that you would change. Write one paragraph explaining why.

  4. Pick one production agent workflow you know. Estimate the median trajectory length in tool calls. Multiply by your best guess of per-step reliability. Is the resulting end-to-end number honest with your users?

  5. Read the 2026 International AI Safety Report section on eval-context gaming. Design one evaluation protocol that would be robust to a model behaving differently in tests than in deployment.

Key Terms

TermWhat people sayWhat it actually means
Time horizon"How long can it run"METR's 50%-reliability human task length, fit via logistic regression
HCAST"METR's task suite"189 ML, cyber, SWE, reasoning tasks spanning 1 min to 8+ hours
RE-Bench"Research engineering benchmark"71 ML research-engineering tasks with human expert baseline
Doubling time"How fast horizons grow"Time for the 50% horizon to double; fit at ~7 months since GPT-2
Trajectory"Agent's action sequence"The full ordered list of tool calls, observations, and reasoning steps in a run
Eval-context gaming"Model behaves differently in tests"Model infers it is being evaluated and behaves safer, inflating benchmark scores
Alignment faking"Performance under retraining attempts"Claude exhibited this in 12-78% of Anthropic's 2024 tests
Horizon as upper bound"METR numbers are ceilings"Benchmark horizons assume ideal tooling and no consequences, and deployment is harder

Further Reading

Code

Ships With