Orchard: An Open-Source Agentic Modeling Framework

Motivation and design

• The framework addresses infrastructure bottlenecks caused by closed or rigidly coupled environment layers, which limit reproducibility, reuse, and independent comparison of data, recipes, and evaluations.Orchard instead supports reuse across task domains, agent harnesses, and stages including trajectory distillation, on-policy RL, and evaluation.Source: §1 · pp. 2–3

Orchard Env

• Orchard Env is a thin, harness-agnostic Kubernetes-native service exposing reusable sandbox, execution, file, network, and REST primitives without coupling to trainers or task domains.Runtime agent injection, direct Pod-IP routing, asynchronous lifecycle management, heartbeat cleanup, and watch-based readiness tracking support scalable interaction; command latency averages 0.28 seconds, with 100% success in a 1,000-sandbox stress test.Source: §1 · p. 3, Figure 2 · p. 3

Framework scope

• Across domains, the same environment abstraction supports heterogeneous modalities, tool interfaces, harnesses, and reward mechanisms while enabling reusable SFT, RL, trajectory, and evaluation artifacts.The released framework includes the environment service, training recipes, and trajectory datasets for software engineering, GUI navigation, and personal-assistant tool use.Source: §1 · pp. 3–4, Figure 2 · p. 3

Orchard-SWE

• 73.0% on SWE-bench Verified makes Orchard-SWE state of the art among comparable-size open-source models, while approaching frontier systems more than 10× larger.RPR-based RL reaches 69.7%, and historical experience distillation enables value-model reranking to reach 73.0%.Source: Figure 1 · p. 2, §1 · p. 4

Orchard-GUI

• 68.4% average success across WebVoyager, Online-Mind2Web, and DeepShop makes Orchard-GUI the strongest open-source GUI agent despite its 4B backbone.Its benchmark success rates are 74.1%, 67.0%, and 64.0%, respectively, and it remains competitive with proprietary systems.Source: Figure 1 · p. 2, §1 · p. 4

Orchard-Claw

• Orchard-Claw reaches 59.6% pass@3 on Claw-Eval with 0.2K synthetic tasks, rising to 73.9% with the stronger ZeroClaw harness at inference.The corresponding pass3 results are 31.7% and 41.0%, respectively, demonstrating transfer across harnesses.Source: §1 · p. 4

2 Orchard Env

• Orchard Env defines a thin standalone service boundary independent of agent harnesses, trainers, rollout abstractions, and inference backends.It is designed to support heterogeneous task environments and arbitrary Docker images with low adaptation cost on standard cloud infrastructure.Source: §2 · pp. 4–5, §2.2 · p. 6

2.1 Architecture Overview

• Orchard Env uses a three-layer architecture: synchronous/asynchronous Python clients, a Kubernetes orchestrator, and a runtime-injected in-pod agent.The orchestrator handles lifecycle and scheduling, while execution traffic goes directly to sandbox agents; runtime injection supports arbitrary user-supplied images without rebuilding them.Source: §2.1 · pp. 5–6

2.2 Comparison with Existing Systems

• Runtime agent injection lets Orchard Env support hundreds of heterogeneous task environments, including the diverse images required by SWE-bench, without per-image modifications.The full stack runs on managed or self-hosted Kubernetes with researcher control over resources, networking, and autoscaling.Source: §2.2 · p. 7
• $673 with spot instances makes Orchard Env 10× cheaper than managed alternatives, while its $3,362 on-demand cost is below Daytona and E2B at $7,078 each.The comparison covers 128 parallel 2-vCPU, 8-GiB sandboxes over 240 hours.Source: §2.2 · p. 7, Table 2 · p. 8

2.3 System Evaluation

• 0.28 s average execution latency matches SkyPilot Code Sandbox (0.284 s) and beats E2B (0.747 s) and Modal (2.046 s).Direct Pod-IP communication bypasses the Kubernetes API server on the hot path.Source: Table 3 · p. 8, §2.3 · p. 8
• Orchard Env matched direct Docker across three Terminal-Bench 2.0 models, with a marginal 1–2-point edge in every case.Numbers were averaged over three independent runs per model-backend pair, indicating no observable execution overhead or interference.Source: Table 3 · p. 8, §2.3 · p. 9
• 100% of 1,000 parallel sandbox sessions succeeded through creation, four executions, and deletion in 26 seconds.This corresponds to roughly 154 commands per second and relies on watch-based readiness tracking, per-sandbox locking, and heartbeat cleanup.Source: Table 4 · p. 8, §2.3 · p. 8

3 Orchard-SWE

• Orchard-SWE uses Orchard Env for isolated execution while training agents through SFT on teacher trajectories followed by RL with environment-grounded rewards.Evaluation targets SWE-bench Verified, where correctness requires passing the complete gold test suite; auxiliary benchmarks include SWE-bench Multilingual and Terminal-Bench 2.0.Source: §3 · p. 9, §3.1 · p. 9

3.2 Trajectory Collection and Curation

• The released corpus contains 107K trajectories across 19,287 tasks, including 74.6K resolved and 32.5K unresolved trajectories.Both trajectory types are retained: resolved traces provide imitation signal, while unresolved traces support partial-progress supervision after filtering and curation.Source: §3.2 · p. 10, Table 5 · p. 10

3.3 Off-policy Learning: Supervised Fine-Tuning with credit assignment

• Credit-assignment SFT converts unresolved trajectories into supervision by estimating outcome-conditioned per-step values and training on positive temporal-difference rise segments.The resulting value curve is inverted-U in 98.9% of annotated trajectories, while resolved trajectories retain all action tokens as training targets.Source: §3.3 · p. 11

3.4.2 Denser Reward: On Policy Distillation (OPD)

• On-policy distillation (OPD) supplies token-level dense rewards from a stronger teacher, complementing sparse task outcomes with step-by-step distribution-matching supervision.The section extends OPD beyond reverse KL to general f-divergences, whose induced reward enters a REINFORCE-style score-function gradient.Source: §3.4 · p. 12, §3.4.2 · pp. 14–15

3.4.1 Reinforcement Learning with Balanced Adaptive Rollout (BAR)

• Balanced Adaptive Rollout (BAR) assembles fixed-size per-prompt groups online until their positive-reward fraction reaches a target interval, reducing zero-variance waste and imbalance noise.It adapts generation to prompt difficulty, supports group-relative estimators, and falls back to relaxed selection when the maximum rollout budget is exhausted.Source: §3.4.1 · pp. 12–13
• BAR combines online group assembly with final group filtering to produce information-dense, fault-tolerant RL batches despite sandbox failures, truncation, and unusable trajectories.The filter preserves gradient quality by replenishing dropped groups from oversampled prompts, while BAR maximizes the chance that groups satisfy the filter initially.Source: §3.4.1 · pp. 13–14

3.4.2 Denser Reward: On Policy Distillation (OPD)

• SWE-bench Verified solve rate rises from about 60% to 70.3%, while mean LLM-as-a-Judge reward increases to 0.76 over RL training.These trends track outcome and process rewards, respectively, during the reported training run.Source: Figure 4 · p. 16
• JS-induced OPD slightly outperforms reverse-KL variants while producing fewer gradient-norm spikes and healthier policy entropy during training.Combining GRPO with OPD JS improves performance further, matches GRPO+BAR, and reduces rollout costs when BAR is omitted.Source: Table 7 · p. 15, §3.4.2 · p. 16

3.4.3 Denser Reward: Rubric-based Process Reward (RPR)

• RPR scores entire trajectories with rproc(τ) ∈ [0, 1], rewarding tests, reproduction, assertions, and failure handling that binary outcome rewards cannot capture.The judge evaluates process quality as independently as possible from whether tests ultimately pass.Source: §3.4.3 · p. 16
• The rubric scores seven dimensions, with R1 assessing the fix and R2–R7 assessing self-verification through tests, reproduction scripts, and checks.Scores are independently reasoned from concrete commands and outputs rather than holistic impressions.Source: §3.4.3 · pp. 16–17
• Self-verification receives 0.85 of the aggregation weight, with test relevance weighted most heavily because reproducing the issue provides the strongest evidence of correctness.The seven scores are combined by a fixed weighted average.Source: §3.4.3 · p. 17
• When ground truth exists, a held-out [REFERENCE] block grounds self-test relevance, coverage, and discriminativeness without requiring exact gold-edit matching or hidden-test execution.Reference guidance applies to R2, R3, and R5.Source: §3.4.3 · p. 17
• RPR penalizes responses to evidence—such as ignoring failures or mislabeling outputs—rather than error strings themselves, keeping process scores decoupled from terminal outcomes.Each trajectory is serialized into a prompt containing the issue, optional reference, trace, and patch; the judge returns evidence-citing rationale and JSON dimension scores.Source: §3.4.3 · p. 17

3.5 Distilling Historical Rollouts into a Value Model

• Historical experience distillation trains a standalone Qwen3-4B value model offline on heterogeneous, outcome-verified trajectories to score unseen-policy rollouts across distribution shift.The model uses turn-level value targets, terminal rewards propagated with γ = 0.98 and λ = 0.95, and supports inference-time best-of-N selection without new environment interaction.Source: §3.5 · pp. 17–18

3.6 Main Results

• 73.0% resolve rate makes Orchard-SWE the strongest open-source method, approaching frontier systems over 10× larger through offline historical-value-model reranking.Reranking the 69.7% RPR policy with a 4B value model at best-of-N, N = 6, adds 3.3 points; an oracle reaches 82.4%.Source: Table 8 · p. 19, §3.6 · p. 20
• 69.7% is Orchard-SWE’s best single-model result with RPR, exceeding BAR’s 69.1% and OPD’s 69.3% on Qwen3.5-35B-A3B.RPR rewards verification discipline as well as terminal success, improving both solve rate and self-verification behavior.Source: §3.6 · p. 20

3.7 Generalization to Unseen Harnesses and Tasks

• Orchard-SWE retains 45.0–64.3% across three harnesses, while single-harness baselines suffer malformed outputs or drops as large as 58.8 points.It also achieves 51.0% on SWE-bench Multilingual and 20.1% on Terminal-Bench 2.0, versus 28.7% and 0.0% for OpenSWE-32B.Source: §3.7 · pp. 20–21

3.8 Ablations and Analysis

• Data scale dominates selection strategy: increasing Diverse-repo training from 512 to 2048 trajectories raises resolve rate from 44.0% to 52.2%, while strategy spread shrinks from 5.5 to 2.0 points.Property-based selectors help at small scale but converge with heuristic baselines as data increases.Source: Table 10 · p. 22, §3.8 · p. 22
• Cross-harness ablation shows matched-training evaluation reaches 53.5–57.9%, whereas mismatched evaluation collapses to 19.0–28.0%, demonstrating that multi-harness training is necessary.OpenHands-collected trajectories transfer slightly better to mini-swe-agent than the reverse, but tool formats and observation conventions remain tightly coupled.Source: Table 11 · p. 23, §3.8 · p. 23
• Credit-assignment SFT improves a scale-matched resolved-only baseline from 59.3% to 61.2%, a 1.9-point gain from unresolved rise segments.The full recipe compounds this signal with larger data to reach 64.3% before RL.Source: §3.8 · p. 23
• RL improves moderate SFT from 48.1% to 50.1% on Verified and 22.0% to 28.7% on Multilingual, but heavy SFT improves Verified from 64.3% to 67.5% while slightly reducing OOD transfer.The results indicate a specialization tradeoff: heavy SFT sharpens in-distribution behavior, whereas moderate SFT preserves broader behavioral diversity.Source: §3.8 · p. 23

4.2 Generic Tool-Calling Agent Harness

• Orchard-GUI uses a generic multi-turn ReAct-style harness rather than a bespoke browser-agent framework, separating harness effects from data and training differences while supporting cross-domain agentic learning.The harness exposes 13 atomic OpenAI-format tools, with done(response) as the sole episode-termination mechanism and carrier of the final user-facing output.Source: §4.2 · p. 24

4.3 Trajectory Collection and Curation

• The dataset pipeline filters 292,092 raw WebGym tasks to 15,601 deduplicated seed intents, then samples four teacher rollouts per task and curates a 412-task, 70-website SFT corpus.RL uses a separately deduplicated pool of 2,198 tasks; the seed pool spans 13,063 hosts and covers 85.0% of MOZ Top-500 and 57.0% of SimilarWeb Top-100 websites.Source: §4.3 · pp. 25, 27, Figure 5 · p. 26

4.4 Training Recipe

• The training recipe first applies SFT to teacher-distilled trajectories, then multi-turn GRPO-style RL with format and GPT-4.1 judge rewards in Orchard Env.RL uses dynamic sampling, removes judge failures and captcha-aborted runs, and raises the episode budget from 15 to 30 steps to extend capability to harder tasks.Source: §4.4 · pp. 27–28

4.5 Main Results

• RL initialized from SFT consistently outperforms base-model initialization, ultimately exceeding 50% evaluation success versus below 40% while producing more stable optimization behavior.Although both settings achieve comparable training rewards, supervised initialization provides a stronger behavioral prior for exploration and generalization.Source: Figure 8 · p. 28, §4.5 · p. 29
• 68.4% average success—74.1% on WebVoyager, 67.0% on Online-Mind2Web, and 64.0% on DeepShop—makes Orchard-GUI the strongest open-source result and competitive with Gemini computer-use-preview at 69.3%.The model uses a 4B backbone and only 2.6k training tasks; RL improves the SFT checkpoint by 13.9, 20.0, and 15.3 points across the three benchmarks.Source: Table 14 · p. 29
• On Online-Mind2Web and DeepShop, Orchard-GUI exceeds the strongest prior open baseline by 31.7 and 21.7 points and surpasses its 235B teacher by 3.3 and 7.3 points.The largest gains occur on Online-Mind2Web, whose broad, diverse website distribution tests generalization to unseen interfaces more strongly than the narrower benchmark distributions.Source: §4.5 · p. 29

5.1 Problem Setting

• Claw-Eval scores daily workflow agents by combining safety, completion, and robustness, counting a task as passed when task_score ≥0.75.The evaluation audits complete agent trajectories using automated scripts and LLM-as-a-judge.Source: §5.1 · p. 30

5.2 Trajectory Collection and Curation

• 192 synthesized tasks yield 561 successful trajectories and 4,537 training pairs from five MiniMax-M2.5 rollouts per task across ReAct-style and ZeroClaw harnesses.Tasks are synthesized from Claw-Eval and ClawHub seeds through a four-step proposal, environment-generation, solving, and refinement loop.Source: §5.2 · p. 30

5.3 Training Recipe

• Orchard-Claw uses SFT on teacher-distilled trajectories followed by binary-reward GRPO RL, with Orchard Env running 64 asynchronous rollout sandboxes per step.The model uses Qwen3-30B-A3B-Thinking-2507; RL uses a 10-minute wall-clock budget rather than a maximum turn limit.Source: §5.3 · p. 31

5.4 Main Results

• 59.6% pass@3 on Claw-Eval lets Orchard-Claw surpass its backbone and similarly sized specialized models after training on only 0.2k synthetic tasks.RL adds 9.3 absolute points to both pass3 and pass@3 over the SFT checkpoint.Source: §5.4 · p. 31, Table 15 · p. 32
• 73.9% pass@3 with ZeroClaw, versus 59.6% under the native ReAct-style harness, gives Orchard-Claw the largest harness-switching gain among compared models.The corresponding pass3 scores are 41.0% and 31.7%, respectively; the improvements are +14.3 and +9.3 absolute points for pass@3 and pass3.Source: Table 16 · p. 32, §5.4 · p. 32

Interactive environment orchestration

• Orchard Env exposes a decoupled REST service for trajectory distillation, on-policy rollouts, and evaluation, unlike integrated or harness-bound alternatives.MegaFlow (Zhang et al., 2026b) co-designs Model, Agent, and Environment services, while ProRL Agent separates rollouts but retains AgentHandler-bound environments.Source: §6 · pp. 32–33
• Orchard Env targets open-source RL research with tunable resource limits, heartbeat lifecycle control, network isolation, and lower overhead than commercial sandbox platforms.E2B (E2B, 2024), Daytona, and Modal provide sandbox APIs but generally lack these fine-grained training controls.Source: §6 · p. 33

Software engineering agents

• Orchard-SWE scales trajectory quality through frontier-model distillation and partial-credit supervision on failed traces rather than synthesizing new software-engineering tasks.SWE-bench and SWE-bench Verified define its benchmark setting, while mini-swe-agent and OpenHands provide its two training harnesses.Source: §6 · p. 33

GUI and browser navigation agents

• Orchard-GUI evaluates a unified model on WebVoyager, Online-Mind2Web, and DeepShop, spanning live, structurally distinct web tasks with non-overlapping action spaces and reward signals.These benchmarks test cross-domain generalization without benchmark-specific tuning; prior work progresses from prompting and SFT toward RL-based grounding and reasoning.Source: §6 · pp. 33–34

Generalist long-running autonomous agents (Claw-agent)

• Claw-style agents maintain persistent state, identity, skills, and proactive workflows, making cross-harness generalization difficult as their heterogeneous tool surfaces expand.Claw-Eval, ClawGym, MetaClaw, and OpenClaw-RL represent recent evaluation and training directions.Source: §6 · p. 34

Orchard Env

• Orchard Env decouples sandbox management from agent harnesses, trainers, and task domains to make trajectory collection, SFT, RL rollouts, and evaluation more reusable, reproducible, and cost-effective.The environment service is designed as a shared infrastructure layer for scalable agentic modeling.Source: §7 · p. 34

Cross-domain recipes

• Orchard-SWE, Orchard-GUI, and Orchard-Claw show that one environment layer can support diverse agents and training recipes across software engineering, GUI navigation, and personal-assistant workflows.The framework also shows improved transfer across harnesses, domains, and pipeline stages.Source: §7 · p. 34

Implications

• By making its environment service, training recipes, and data collection reusable across domains and harnesses, Orchard lowers the barrier to open, reproducible, capability-focused agentic-AI research.The conclusion emphasizes that scalable agentic progress depends on both infrastructure and training design.Source: §7 · pp. 34–35

Design principles

• Orchard Env combines five design choices to remain reusable across training recipes, agent harnesses, and model backends while supporting isolation, lifecycle management, and large-scale training.Agent injection and direct Pod-IP communication are the distinguishing choices; network isolation, asynchronous heartbeat cleanup, watch-based readiness, and Kubernetes-native deployment provide production-grade operations.Source: A Orchard Env Design Details · p. 43

Agent Injection via Init Containers

• Agent injection copies a self-contained Python runtime and agent server through an init container, avoiding modifications to heterogeneous task images.The injected agent runs from a shared emptyDir volume via /opt/sandbox-agent/start.sh, targeting Linux container images by default.Source: A Orchard Env Design Details · p. 43

Direct Pod-IP Communication

• Direct Pod-IP communication routes execution and file operations around the Kubernetes API server, reducing command overhead and removing a high-concurrency throughput bottleneck.This also avoids Kubernetes exec API control-plane mediation and WebSocket setup overhead.Source: A Orchard Env Design Details · p. 43

Network Isolation

• NetworkPolicy-based isolation blocks outbound sandbox connections by default while allowing selectively enabled per-sandbox egress that is removed with the sandbox.This provides defense-in-depth against data exfiltration by user-supplied commands.Source: A Orchard Env Design Details · p. 43

Asynchronous Lifecycle with Heartbeat-Based Cleanup

• Asynchronous sandbox creation returns immediately, while heartbeat expiry triggers background deletion to prevent resource leakage from crashed or abandoned clients.Clients poll or block on /wait until readiness, and periodic heartbeat messages keep long-running sandboxes alive.Source: A Orchard Env Design Details · p. 43

Watch-Based Readiness

• A persistent LIST+WATCH stream tracks pod state transitions in real time, caches them in memory, and notifies waiters without repeated Kubernetes API polling.Waiters are notified through asyncio.Event.Source: A Orchard Env Design Details · p. 43

Orchard setup

• Orchard runs 128 sandboxes on 16 spot-backed D16ads_v5 nodes, with a separate standard-priced node orchestrating the deployment.Each D16ads_v5 node provides 16 vCPUs and 64 GiB RAM and hosts eight sandboxes.Source: B Cost Analysis Details · p. 44

Spot-instance economics

• Spot pricing reduces D16ads_v5 costs to approximately $0.165/hour versus $0.824/hour on demand, and ephemeral sandboxes can be recreated after preemption.Managed services cannot pass through spot pricing because they control the underlying infrastructure.Source: B Cost Analysis Details · p. 44

Managed-service comparison

• Packing eight sandboxes per VM shares overhead, whereas MegaFlow assigns one full instance per task; managed-service estimates are $0.2304/hour for E2B and Daytona and $0.3354/hour for Modal.MegaFlow’s estimated allocation exceeds the 2-vCPU target, while Modal Sandbox is nonpreemptible by default.Source: B Cost Analysis Details · p. 44

Cost comparison

• $3,362 on demand is roughly half E2B and Daytona ($7,078 each) and one-third of Modal ($10,305), while giving researchers full cluster control.The comparison estimates 128 parallel 2-vCPU, 8-GiB sandboxes running for 240 hours.Source: B Cost Analysis Details · p. 44

Research-scale implications

• These savings compound across thousands of environment-interaction hours needed for 160K rollout trajectories, ablations, and training-recipe iterations.The appendix argues that Orchard’s self-hosted, spot-friendly design makes such workloads practical for academic research budgets.Source: B Cost Analysis Details · p. 44

System Design and Component Orchestration

• The RL system extends slime with four loosely coupled Ray- and HTTP-connected services that can be independently scaled, replaced, or restarted.The services comprise the policy trainer, rollout inference service, sandboxed execution service, and agentic loop driver.Source: C Orchard RL Implementation Details · p. 45
• The agentic loop asynchronously coordinates tool-calling by querying the latest policy, executing structured bash calls in isolated sandboxes, appending observations, and enforcing step, wall-clock, and token budgets.The inference service supports KV-cache reuse, deterministic sampling seeds, and per-token log-probability extraction.Source: C Orchard RL Implementation Details · p. 45
• Pipelined rollouts overlap generation with weight updates, while centralized management tolerates partially failed trajectories so individual sandbox failures do not crash optimization.The rollout manager dispatches rollout k+1 while the trainer updates weights for rollout k.Source: C Orchard RL Implementation Details · p. 45
• The implementation bounds tail latency through layered timeouts, retries, resource escalation after exhaustion, jittered sandbox creation, and token-space loss masking that excludes tool-result tokens from gradients.Orchard Env’s 0.28 s command latency is critical because each rollout requires dozens of environment interactions and throughput scales with sandbox responsiveness.Source: C Orchard RL Implementation Details · pp. 44–45

Credit assignment

• Credit assignment estimates P(resolve) at each checkpoint by reasoning backward from the known failure and lowering probability at or before the critical mistake.The prompt requires probabilities to rise during correct exploration, drop at the critical error, and remain low through submission.Source: D Orchard-SWE Prompt List · p. 45
• Calibration uses a 0.3–0.5 base rate, requires a drop after the critical error, and constrains the failed trajectory’s final probability below 0.2.The output is a JSON array with one P(resolve) estimate and reasoning entry per step.Source: D Orchard-SWE Prompt List · pp. 45–46

Process-reward grading

• Process grading scores seven dimensions from 0.0 to 1.0, separating final patch quality from the quality of the agent’s self-verification.Scores are combined by weighted average, must cite concrete trajectory evidence, and should be judged independently of whether tests ultimately passed.Source: D Orchard-SWE Prompt List · p. 46
• The rubric penalizes mishandling evidence rather than incidental error strings, and uses any reference block only to judge self-test relevance, coverage, and discriminativeness.It does not grade exact-match edits or penalize agents for omitting hidden official tests.Source: D Orchard-SWE Prompt List · p. 46
• Self-verification evaluates whether tests target the issue, cover relevant cases, use precise assertions, reject incomplete fixes, guard regressions, and preserve verification integrity.A genuine before-and-after reproduction is strongest for relevance, while ignored failures or submission without verification receive the lowest integrity score.Source: D Orchard-SWE Prompt List · p. 47

• RPR assigns 0.85 of its weight to reference-guided self-verification criteria R2–R7, with R1 grading patch quality.The rubric weights sum to 1.Source: Table 17 · p. 48
• All tracked LLM-as-a-Judge process-reward metrics, including fix quality, test relevance, coverage, regression guarding, and verification integrity, trend upward during RL training.The figure also reports their mean, which increases with training.Source: Figure 10 · p. 48

Tool-schema organization

• The appendix defines 13 atomic tools in OpenAI tool-call JSON Schema, grouped by action family for readability.At each step, the agent emits one or multiple tool calls enclosed in <tool_call>...</tool_call> blocks.Source: F Orchard-GUI Tool List · p. 49

Pointer Actions

• Pointer actions support click, hover, and drag operations using viewport-relative coordinates.Click accepts [x, y] coordinates and configurable click counts or mouse buttons; hover moves to coordinates, while drag specifies start and end coordinates with optional intermediate steps.Source: F Orchard-GUI Tool List · p. 49

Keyboard Actions

• Keyboard actions provide write for replacing focused-field contents and press_keys for individual or simultaneous key presses.press_keys supports modifiers, navigation keys, standard alphanumeric keys, and configurable repetition.Source: F Orchard-GUI Tool List · p. 50

Page-Navigation Actions

• Page-navigation actions include scrolling pages or nested elements, direct URL navigation, browser-history back navigation, and timed waits.Scroll supports directional movement, viewport-relative amounts, and optional element coordinates; goto_url opens a specified URL in the current tab, while wait allows loading or animations to complete.Source: F Orchard-GUI Tool List · p. 51

Tab-Management Actions

• Tab-management actions open, switch, and close browser tabs, while done signals task completion or failure with a final response.switch_tab uses a 0-based tab index, and done can report successful completion, inability to proceed, or CAPTCHA and anti-bot blocking.Source: F Orchard-GUI Tool List · p. 52

• The rollout shows seven complete think→tool_call→tool_response cycles, with the final-step prompt containing the accumulated trajectory and one context image.The displayed response ends with reasoning followed by a terminating done call, while the JSON tool schema is abbreviated in this appendix.Source: G Example GUI Agent Trajectory · p. 53
• The agent receives action history, the user request, and page observations including tabs, screenshots, and an optional accessibility tree before acting.Its output must begin with <think> and include executable <tool_call> blocks; multiple calls execute sequentially when appropriate.Source: G Example GUI Agent Trajectory · p. 53
• The agent applies Amazon’s “Washable” filter and sees 464 results for dog beds.The request is for a washable dog bed at least 30 inches long.Source: G Example GUI Agent Trajectory · pp. 54, 56
• The agent identifies the EHEYCIGA Washable Large Dog Bed as a candidate because its stated dimensions are 35x23 inches, satisfying the at-least-30-inch length requirement.The bed is already covered by the Washable filter, and the agent clicks it to verify the product details.Source: G Example GUI Agent Trajectory · pp. 56–57
• 35 inches: The agent found a washable EHEYCIGA dog bed measuring 35" x 23" x 3.5", exceeding the 30-inch minimum.The length was supported by both the product title and page specifications.Source: G Example GUI Agent Trajectory · p. 57
• The agent concluded that the task was complete because the product matched all specified requirements and required no further action.Source: G Example GUI Agent Trajectory · p. 57
• $17.99 and 4.7 stars from 381 reviews: The final response also reported the bed’s price, rating, Amazon’s Choice label, in-stock status, and suitability for dogs up to 45 lbs.Source: G Example GUI Agent Trajectory · p. 57

H Orchard-GUI Task Filtering Pipeline

• Five filtering stages remove benchmark overlap, child-task redundancy, residual WebVoyager intents, long-tail websites, and near-duplicate tasks.The stages target contamination, intra-family redundancy, noisy websites, and semantic duplication.Source: H Orchard-GUI Task Filtering Pipeline · p. 58
• The retained pool keeps complementary PAE-WebVoyager and InSTA-v3 tasks, restricts sites to major popularity lists with at least two tasks per site, and removes intents with cosine similarity above 0.99.Semantic deduplication uses Qwen/Qwen3-Embedding-8B after the benchmark and website filters.Source: H Orchard-GUI Task Filtering Pipeline · p. 58
• 15,601 unique task intents remain after filtering and serve as the seed set for sampling teacher trajectories for SFT and RL prompts.The pipeline starts from 292,092 raw WebGym task instances.Source: H Orchard-GUI Task Filtering Pipeline · pp. 57–58

Example Model Input in the ZeroClaw Harness

• The ZeroClaw system prompt instructs the assistant to use tools directly, remain concise and helpful, and format function calls as JSON within XML tags.The available interface includes function signatures and requires each call to specify a function name and arguments object.Source: I Example Claw Agent Trajectory · p. 58
• The user assigns a multi-step inventory task involving email review, inventory checks, supplier contact, weekly follow-ups, and writing a rotation-plan file.The request specifically asks the agent to read Diana’s email and Tom’s follow-up before completing the operational actions.Source: I Example Claw Agent Trajectory · p. 58

Example Response from LLM

• The model decomposes the request into checking critical and expiring items, grouping supplier quote emails, scheduling critical-item follow-ups, writing the summary, and flagging items above $1000.It interprets “expiring soon” as within 45 days of April 10, 2026, before May 25, 2026.Source: I Example Claw Agent Trajectory · p. 59
• The agent begins execution by issuing an inventory_list_products function call after stating that it will retrieve the inventory list.This is the first concrete tool action shown in the example response.Source: I Example Claw Agent Trajectory · p. 59