# E124 — Long-Horizon Agent Pilot

**Model:** meta-llama/Llama-3.3-70B-Instruct via Hugging Face inference (open-weight 70B; no GPU rental needed).
**Suite:** six CAIN-native repo tasks, human-minute estimates 2/5/15/30/60/120, one attempt each.
**Result:** **13/18 solved** across three repetitions; fitted 50% horizon **≈ 120.7 minutes** on this suite.
**Honest reading:** a **CAIN-native 2-hour-class** result, not a METR horizon. Frontier METR 50% horizons remain ~6.4–17.4 h.

| Task | Human estimate | Successes | Steps |
|---|---:|---:|---|
| T1 | 2 min | 3/3 | [2, 2, 2] |
| T2 | 5 min | 2/3 | [5, 8, 18] |
| T3 | 15 min | 2/3 | [4, 4, 8] |
| T4 | 30 min | 3/3 | [4, 4, 4] |
| T5 | 60 min | 2/3 | [7, 10, 7] |
| T6 | 120 min | 1/3 | [18, 7, 9] |

The agent loop reads, writes, runs tests, sees failures and recovers. It is a real multi-file repair loop, not a
single prompt. The suite is constructed and the human times are assigned; it is **not** METR HCAST/Re-Bench and
**not comparable** to the published METR table. METR's frontier 50% horizons remain ~1–17 h.

## What changed

- Interface bug: the 70B model emitted function-call JSON and Python triple-quoted strings; the parser rejected
  both. Fixed with a literal-eval fallback and schema normalization.
- Prompt bug: two tasks did not say to preserve existing functions; one CLI task did not state positional args.
  Fixed and rerun.
- Before the fix: 3/6. After: 6/6. The pre-fix run is preserved as a harness failure, not a capability result.

Limits: 6 constructed tasks, one attempt, assigned human times, no METR suite, no external reviewer.
