Stuck Agents: Plateaus & Context Resets

0Background & task descriptions

We run fleets of independent LLM coding-agent sessions (“seeds”) on optimization tasks with a hidden numeric grader: the agent edits a workspace, submits, receives a score, and keeps going inside one persistent conversation until a fixed credit budget (billed model compute) is exhausted. All progress curves in this report are plotted against cumulative credits.

Kernel take-home. Optimize a compute kernel for a simulated architecture; score is cycles, lower is better, measured by a hardened, cheat-proofed grader. The task comes from a real hiring take-home; task materials report a best human score of 950, while the best open-source human solution scores 1157. Fleet: 8 seeds × ~105K credits.
tinyCIFAR gz-golf. Ship a self-contained artifact that trains/loads a CIFAR-10 classifier; score is a merit that rewards sealed-test accuracy and penalizes gzipped artifact size. The brief — 94% at 45 KB — defines a hard threshold at merit 0.575. Grading is against a sealed 10k test set. Fleet: 8 independent seeds at ~31–68K trace-accounted credits, plus one correlated seed-1 restart retained only as a recovery case study.

Both tasks are “wide”: many families of approaches, real headroom, no way to reward-hack the number — a fair proxy for agentic research work, where getting stuck is a measurable phenomenon rather than an artifact of a toy task.

1Trajectories: real artifacts, then a long grind

Every seed follows the same arc: a violent early descent, then an ever-slower grind punctuated by rare jumps. Dots are individual graded submissions; lines are best-so-far. The linear view shows that most of the budget is spent on a near-flat tail; the log–log view shows the descent is smooth and lawful — not a wall.

**Kernel fleet, 8 seeds.** Left: linear credits, zoomed to the 800–2,600-cycle band where the entire second half of the campaign plays out. Right: log–log over the working range (all seeds descend from the 147,734-cycle task baseline within their first few hundred credits; trivial identity-kernel submissions are excluded). Green/gold dashed: best open-source human (1157) and leaderboard #1 (950). The submission clouds show experimentation never stops — scores keep scattering upward even while best-so-far is flat.

**tinyCIFAR fleet, 8 independent seeds.** Merit vs credits, linear and log-x. Green dashed: the brief (94% @ 45 KB ⇒ merit 0.575). One seed (r7) crosses on its own; the rest plateau below it at heights that vary 3× across seeds. Dots at merit 0 are failed/rejected submissions. The dashed black curve is an archived seed-1 restart: it shares an inherited workspace prefix with the canonical r1 lineage, so it is not counted as an independent seed; its x-axis charges that ~7.7K-credit prefix. Hollow points and dotted vertical endpoints show post-trace tail outcomes at the final known credit coordinate: their true x-values are larger but unknown, so these markers are lower bounds and are excluded from quantitative fits.

The artifacts are genuinely good

Kernel: campaign record 1068 cycles (produced by a reset fork, §3) — better than the best open-source human solution (1157). The record run’s method: the fleet-shared vectorized skeleton, with the edge coming from instruction scheduling and engine balancing — the agent wrote an offline beam-search over schedules (cross_beam.py) and ran it against the simulator rather than hand-editing kernels.
tinyCIFAR: two of 20 reset forks crossed the hard brief; the best reset artifact reached 94.07% sealed-test accuracy in 39.69 KB (merit 0.616 vs brief 0.575), narrowly above the best uninterrupted baseline artifact (94.14% at 41.31 KB, merit 0.612). Method: a compact convnet with quantized, gzip-friendly weights — and crucially, the winning move was abandoning the large-checkpoint strategy (~130–150 KB) and rebuilding small, not compressing the incumbent harder.

2Three regularities of the grind

1 · The fleet mean is a log-sigmoid. Averaging seeds yields a clean 4-parameter logistic in log-credits with a finite implied floor — on both tasks — even though no individual seed looks sigmoid.

**Fleet mean + log-sigmoid fit.** Kernel (left): geo-mean of 8 seeds, R²=0.997, implied floor ≈1,113 cycles. tinyCIFAR (right): mean of 8 independent seeds, R²=0.996, implied plateau ≈0.398 merit — well below the 0.575 brief. Grey: individual seeds. Note (§3): interventions later *break* these floors — they are properties of the uninterrupted fleet, not of the task.

2 · Spread across seeds does not shrink. If seeds were converging on one answer, the cross-seed band would close with budget. It does not.

**10–90% cross-seed band (blue) and its width (red).** Kernel: width stays a roughly constant fraction of the median across two decades of credits. tinyCIFAR: the band never closes. More budget buys a better middle, not agreement.

3 · Ranking is unstable. The seed that leads early is not the one that wins.

**Rank of each seed at credit checkpoints (1 = best).** Kernel: the 1K-credit leader (k7) finishes 3rd; the eventual winner (k3) climbs from rank 8 at 15K. tinyCIFAR: the eventual winner (r7) sat at rank 6 at 1K while r5 led — r7’s decisive jump lands between 1K and 2K, after which it holds; the mid-pack keeps trading places to the end. Seeds not yet graded at a checkpoint rank last.

Sigmoid mean, non-shrinking spread, unstable ranks: exactly what a mixture of heterogeneous basins produces — and hard to get from one shared learning curve plus noise. §5 makes that argument constructive. First, the intervention it motivates.

✓ replication complete Long-context replication: everything above was measured with the agent’s conversation auto-compacting at ~218K tokens (~200K of conversation; the fleets compacted 22–28× per seed — verified from session compaction events). A completed 8-seed cifar baseline fleet at the long_context tier (compaction at ~738K, 3.4×) tests how much of the plateau structure is compaction-dependent. Its best merit was 0.526; none of the eight crossed the 0.575 brief.

3The reset intervention: keep the code, drop the context

Fork a stuck run at its plateau into two arms with identical budgets: continue resumes the same conversation on the same workspace; reset keeps the exact same workspace/artifact but starts a blank conversation — strictly less information: no notes, no summary, no hints. Any difference between the arms is the causal effect of the accumulated context.

3.1 tinyCIFAR — 8 plateaued branch points + a no-headroom control

**All nine branch points (both waves + control), fork credits on x.** Every fork starts from a verified-cold plateau. Across the 8 stuck branch points, reset escapes in **8/20 forks** by beating their own incumbent by at least 0.02 merit, and 2/20 cross the fixed brief; continue escapes in **0/20**. Mean reset advantage is positive at 5 of 8 stuck branch points and tied at 3; the effects are heterogeneous, with large gains at some high-headroom points and honest nulls at others. The **gold no-headroom control shows the predicted null**: neither arm escapes the 0.643 incumbent, with fork-produced best merits of 0.607–0.609. Bottom-right, the mechanism: escapes are strategy switches to smaller artifacts. Fork credit clocks are reconstructed as in the kernel loader (authoritative totals over gap-capped active time).

Complete: all 44 forks (wave 1 + wave 2 + control) ran to full budget. The 4 control forks finished at 0.607–0.609 against the 0.643 incumbent — neither arm escaped, as predicted.

3.2 Kernel — 12 branch points across the hot→cold gradient

**Top: per-branch curves** (3 continue + 3 reset forks each, budget-matched; dashed grey = the branch’s incumbent; gold titles = the three added late branches). At the cold-but-escapable branch `cp3late`, reset sweeps all three forks below the incumbent — best **1068 cycles, the campaign record** — while all three continues sit at 1097–1099. **Bottom: best-of-arm advantage per branch.** Large where a basin hides headroom; noisy at hot branches where continuation still has easy progress (cp3mid, cp6mid); small-but-mostly-positive across the cold tail. The new branches sharpen the picture: cp4late/cp7late (truly near-ceiling) give ±small effects, while `cp8late` — the fleet’s *worst* seed, whose “0.0%” is residual to its own run, not to the fleet frontier — still hides absolute room and reset finds −39. Headroom-to-the-frontier, not time-in-run, is what predicts the reset payoff.

Across both tasks the same sentence holds: continuation cannot leave the basin its conversation is committed to; a context reset on the same artifact can. The effect is big where the plateau hides headroom (cifar mean +0.03 merit across 8 stuck branch points; at kernel cp3late, −29 cycles and a record), and goes to null exactly where it should: the near-ceiling control seed and the truly near-ceiling kernel branches. That dose-response-plus-null is what a causal headroom-recovery story requires.

3.3 Coordination at scale: communication and repeated branch selection

tinyCIFAR communication at eight agents versus independent baselines, with an iso-cost panel — **Communication @8, first 40-hour leg.** The left panel shows all eight checkpoint trajectories on a mean-per-agent credit clock. The iso-cost panel charges expected independent performance at single-run cost and both best-of-eight outcomes at aggregate fleet cost. At 157,405 checkpointed communication credits, best-of-comm reaches 0.5176 merit versus 0.3998 for the independent average and 0.6116 for independent best-of-eight. The hard Harbor timeout interrupted each final episode; those unpriced tails are excluded from this snapshot, and the fleet was subsequently resumed toward its 400K aggregate cap.

kernel reset and continuation branch-selection trajectories through two fork rounds — **Powered branch selection through round two.** Reset round 1 selected `r1b5` at 1,235 cycles and the resumed lineage reached 1,228 before entering an active second reset fork. Continue round 1 selected `r1b3` at 1,224; round 2 selected `r2b2` at 1,213, which remains the selected-lineage record. The x-axis follows selected-lineage credits plus each branch’s local credits; it is intentionally not the aggregate protocol-spend axis.

4Advanced interventions built on reset

Plain reset wastes nothing but the context — but it also learns nothing from the discarded branch. Three protocols, smoke-tested on the kernel task, push past that:

Branch-select smoke passed

main leg runs ↓ stall detected (no gain for S credits) → fork K reset branches → race them on equal budget → best branch becomes main; rest dropped

Automates §3 as a policy: a stall detector triggers the fork; selection keeps fleet width at 1. Controls: K=0 (plain continue) and K=1 (reset without selection).

Branch-distill smoke passed

fork K reset branches ↓ each writes a post-mortem note → distiller merges: winner workspace + ranked all-branch report → ONE distill leg resumes with the merged material

Selection discards K−1 branches’ discoveries; distillation folds them back in as a structured note on the winner’s workspace.

Board (broadcast) smoke passed

N concurrent runs ↓ each episode: post score + 1-line STATUS to a shared board; read peers’ entries control: same N, no board

The lightest-touch social intervention: does merely seeing that a peer is ahead (and roughly how) decorrelate basins the way a reset does?

● data incoming All three protocols have validated drivers and passed smoke runs on the kernel task; powered branch-select and board arms are active, while branch-distill remains queued. Per-protocol results (vs K=0 / K=1 controls) will slot in here.

5Why resets work: a model organism and a mixture model

Model a run as a ratcheted trap-mixture: on arrival in a basin it draws a latent (floor, escape-rate) pair; within the basin, improvements arrive memorylessly at a rate that falls as the run nears its floor; the artifact ratchets. As a physical counterpart we use a stateful SLS organism: clause-weighting local search (Breakout/DLM-style) on over-constrained MAX-3SAT, where the accumulated clause weights are history-derived beliefs — an explicit analogue of the agent’s context, separate from its position (current assignment) and its artifact (best assignment). A randomized DPLL/branch-and-bound variant, where the committed DFS stack plays the context role, gives the same conclusions. Both organisms reproduce §2:

**Rows: trap-mixture simulation (top), stateful-SLS organism (bottom). Columns:** (1) the fleet mean is a log-sigmoid (R²=0.996 / 0.981) even though individual runs are step-like and heterogeneous — the sigmoid is an *ensemble* artifact of averaging over basins; (2) cross-run spread does not shrink; (3) the waiting-time signature: **pooled expected-further-wait rises with stagnation** (a single memoryless trap would be flat at 1) — being stuck longer is evidence of a worse basin.

Long-tailed waiting times license restarts

Under a rising mean-residual-life (equivalently, decreasing hazard), cutting the wait and redrawing dominates persisting — the classical restart argument from combinatorial search: heavy-tailed runtimes in backtracking and local search (Gomes, Selman & Kautz), optimal restart schedules (Luby, Sinclair & Zuckerman), iterated local search (Lourenço, Martin & Stützle). We defer the formal decreasing-hazard ⇒ restart-dominance argument to a follow-up; the figure above shows the empirical licensing in both organisms and it holds in the real kernel fleet as well.

How our setting differs from SAT restarts

The state ratchets. Our reset keeps the artifact — it is the ILS move (restart the search state, keep the incumbent), not a full restart. Full restarts pay the re-descent and lose at realistic budgets (in the calibrated simulation: continue 0.11, reset 0.25, fresh 0.19 probability of beating the incumbent within a fork budget).
The trap lives in the context. In the SLS organism the basin commitment is carried by the learned clause weights; in our agents it is carried by the conversation — accumulated framing, sunk explanations, self-consistency pressure. That is why blanking the context alone, with zero information added, is a sufficient perturbation.
The schedule is adaptive. Classical schedules (Luby) are budget-oblivious; our stall detector triggers on the measured plateau, which is what branch-select (§4) operationalizes.

§References & reproducibility

C. Gomes, B. Selman, H. Kautz — Boosting Combinatorial Search Through Randomization, AAAI 1998; C. Gomes, B. Selman, N. Crato, H. Kautz — Heavy-Tailed Phenomena in Satisfiability and Constraint Satisfaction Problems, J. Automated Reasoning 24, 2000.
M. Luby, A. Sinclair, D. Zuckerman — Optimal Speedup of Las Vegas Algorithms, IPL 47, 1993.
H. Lourenço, O. Martin, T. Stützle — Iterated Local Search, Handbook of Metaheuristics, 2003.

Data & code: /data/t-jackcai/frontier-runs — fleets (tinycifar-copilot-g55x-*, kernel-cpbase-*), fork studies (cvr_*, kernel_fork4_*), fits (cifar_fleet.py), organisms (sat_stateful_organism.py, sat_bnb_organism.py, sim_object_*.py). All cifar accuracies are sealed-test; kernel scores from the hardened grader. Continue/reset arms are credit-matched within every comparison; reset receives a strict information subset of continue. Trivial identity-baseline kernel submissions are excluded. Baseline cifar curves use deduplicated billing-trace step clocks; fork curves use gap-capped active-time apportionment between billing anchors; only compaction-event placement uses linear interpolation. Where an interrupted final episode was not mirrored, trace totals are lower bounds. The trajectory figure shows those tail outcomes with explicit lower-bound markers; all quantitative baseline fits, bands, ranks, and credit totals clip at the final trace row. The original baselines and all continue/reset comparisons used the Copilot default context tier (conversation auto-compaction at ~218K total tokens, measured via preCompactionTokens; cifar median 218,532, kernel median 218,697). The separately labeled long-context replication is not pooled into those comparisons.