Training Agents to Evolve with Their Harness: TaoLive Digital Avatar Agent Technical Report

TaoLive AIGC LLM Team; Yuhan Sun; Wenhao Lin; Yongdong Luo; Yibo Hu; Meiguang Jin; Junfeng Ma; Weihang Pan

cs.CLAug 16, 2026

Training Agents to Evolve with Their Harness: TaoLive Digital Avatar Agent Technical Report

Authors: TaoLive AIGC LLM Team, Yuhan Sun, Wenhao Lin, Yongdong Luo, Yibo Hu, Meiguang Jin, Junfeng Ma, Weihang Pan, +2 more

Organizations: TaoLive AIGC LLM Team

Abstract

AI-powered digital avatar streamers must answer product questions, engage viewers, and execute marketing strategies in real time, demanding low latency, frequent strategy updates, and accurate yet effective responses. Evolvable Harnesses, whose Skills, Hooks, prompts, and tools can be updated independently of model weights, enable rapid iteration but expose a trade-off: large models adapt zero-shot yet are too slow, whereas compact models meet latency targets but overfit to fixed Harness configurations. We propose Harness-Aware Training (HAT), which trains compact models to adapt to changing Harnesses. Its key component, Harness-State Augmentation (HSA), applies task-preserving transformations to Skill identifiers and content, tool schemas, prompt structures, and Hook functions. Training proceeds in three stages: HSA-SFT learns reasoning and tool use from strong-model trajectories across diverse environments; General On-Policy Distillation restores generalization lost during SFT; and HSA-RL improves robustness to changing Harnesses through reinforcement learning in augmented environments. Across four evaluation sets, HAT achieves 94.8 on Live-Stream QA (base: 80.3; strongest general LLM: 93.0) and 94.6 on Harness-Variant QA (base: 75.4). Unlike Fixed-Harness SFT, which lowers IFEval by 7.7 points from the base model, HAT avoids this regression and reaches 83.5. On one NVIDIA H20 GPU, the optimized system delivers P50 and P95 latencies of 3.4 s and 8.1 s. Deployed in Taobao Live's digital-avatar service, it also yields positive online A/B test results for item-page views.

Figures & tables

Figure 1: Architecture of the live-streaming digital-avatar system and its Harness Agent runtime. (a) Viewer requests and live-room context are processed by the interaction agent. The resulting text response is converted to speech, rendered by the avatar generator, and broadcast to the live room. (b) The expanded runtime operates under an evolvable Harness state

h=(\mathcal{S},\mathcal{T},\mathcal{P},\mathcal{K})

, comprising active Skills, the tool registry, the assembled system prompt, and lifecycle Hooks, respectively. In each agent-loop round, Hooks mediate input processing, model inference, tool execution, and stopping, while the policy may load Skills, invoke externally mapped tools, or propose a final response. A successful stop check emits the response; a failed check returns control to context assembly for the next round.

Figure 2: Harness Evolution with a fixed DeepSeek-V4-flash on dev-set (

n=482

). Optimizing the agent without training model. Evolution 2 is the selected development checkpoint Harness, while later long-tail edits introduce regressions.

Scenario	Proportion
Product Q&A	$\sim$ 46%
Casual chat and engagement	$\sim$ 19%
Clarification follow-up	$\sim$ 16%
After-sales handling	$\sim$ 7%
Discount and promotion inquiry	$\sim$ 4%
Presentation-order adjustment	$\sim$ 2%

Table 1: Approximate intent mix in the available internal traffic summary.

Figure 3: Overview of harness-state augmentation and judging. Left: Examples of augmenting tools, skills, hooks, and prompts. The augmented harness is shared across training stages: a teacher model generates candidate trajectories for SFT, while the policy model produces rollouts for RL. Right: Representative RL trajectories. From top to bottom, Trajectory 1 invokes Tool A, which is unavailable after augmentation, and is penalized by Tool Rationality ; Trajectory 2 selects a nonexistent Skill

\alpha

and is penalized by Skill Selection ; Trajectory 3 violates the modified prompt and is penalized by Accuracy ; and Trajectory 4 correctly adapts to the augmented harness and is rewarded. During SFT, the same judges guide rejection sampling. Together, augmentation and judging teach the model to understand the current harness rather than overfit to a fixed configuration.

Figure 4: Overview of Harness-Aware Training (HAT). Harness-State Augmentation (HSA) transforms the original Harness into diverse, task-preserving states. A strong teacher generates candidate trajectories under these states, which are filtered by a rejection-sampling judge to provide diverse, high-quality supervision for HSA-SFT. Starting from the base model, HAT then follows three stages: HSA-SFT on the filtered trajectories; General OPD, using the base model as the teacher and the HSA-SFT checkpoint as the student to mitigate general-capability degradation; and HSA-RL in changing HSA environments. The resulting policy is the final HAT-trained model.

Set	Size	Source	Evidentiary role
$\mathcal{T}_{1}$ Live-Stream QA	978	Real live-room w/ fixed Harness	Primary industrial quality (live-stream reply).
$\mathcal{T}_{2}$ Harness-Variant QA	978	Real live-room w/ augmented Harness	In-family robustness to new Harness contexts.
$\mathcal{T}_{3}$ Synthetic Live-Stream QA	2023	Synthetic live-room scenarios	In-domain transfer under broader tools/prompts.
$\mathcal{T}_{4}$ IFEval	541	Public Benchmark	General instruction following (official evaluator).
$\mathcal{C}_{\mathrm{judge}}$	482	Real live-room, human-labeled	Judge calibration, not a policy test set. Also dev-set for harness evolving.
$\mathcal{D}_{\mathrm{perf}}$	110	Deployment replay	End-to-end deployment latency.

Table 2: Evaluation sets and sources. The sets cover real live-stream reply quality, in-family Harness variation, synthetic tool and prompt robustness, general instruction following, and actual deployment latency.

Non-augmented
Harness Configuration	$\mathcal{T}_{1}$	$\mathcal{T}_{2}$	$\mathcal{T}_{3}$ Tool Robustness	$\mathcal{T}_{3}$ Prompt Robustness	$\mathcal{T}_{4}$ IFE-P	$\mathcal{T}_{4}$ IFE-I
Base	80.3	75.4	69.5	72.8	81.5	87.7
+SFT	89.5 $\uparrow$ 9.2	88.2 $\uparrow$ 12.8	82.0 $\uparrow$ 12.5	68.2 $\downarrow$ 4.6	73.8 $\downarrow$ 7.7	82.4 $\downarrow$ 5.3
+General OPD	89.5	89.1 $\uparrow$ 0.9	84.3 $\uparrow$ 2.3	72.6 $\uparrow$ 4.4	82.3 $\uparrow$ 8.5	87.9 $\uparrow$ 5.5
+RL	95.1 $\uparrow$ 5.6	94.4 $\uparrow$ 5.3	83.7 $\downarrow$ 0.6	66.7 $\downarrow$ 5.9	82.7 $\uparrow$ 0.4	87.8 $\downarrow$ 0.1
Augmented

Table 4: Training-pipeline ablation. Each row adds one stage to the previous checkpoint;

\uparrow

\downarrow

denotes a descriptive score change from the row above and does not include retraining variance.

Figure 5: Checkpoint trajectories across four training configurations formed by crossing HSA at the SFT and RL stages, evaluated on Accuracy, Effectiveness, Skill Selection, and Tool Rationality. In these trajectories, HSA-SFT is associated with higher reply-quality rewards, while HSA-RL is associated with stronger agentic-behavior rewards, especially Skill Selection and Tool Rationality; combining HSA-SFT and HSA-RL gives the most balanced late-stage reward profile.

Figure 6: Single-run CoT-length diagnostics for the selected HSA-RL trajectory. Faint lines show recorded step values and bold lines show 25-step trailing means. Panel (a) separates tool-call and final-reply CoT lengths and marks the no-penalty (

L\leq 100

), linear (

100<L<200

), and saturated (

L\geq 200

) regions. Panel (b) shows the logged score

q(L)\in[0,1]

and the equivalent raw reward contribution

-0.1q(L)

. The figure is a training diagnostic, not a causal ablation of latency or quality.

Vendor API
Configuration	MTP	$C$	Wall P50 (s)	Wall P95 (s)	TTFT P95 (s)	Decode (tokens/s)	$\leq$ 15s
DeepSeek-V4-flash	–	1	11.210	21.191	2.067	103.44	71%
DeepSeek-V4-flash	–	2	11.976	26.278	1.291	97.01	71%
DeepSeek-V4-pro	–	1	14.312	28.882	1.312	53.53	61%
DeepSeek-V4-pro	–	2	13.752	23.393	1.248	55.04	60%
Qwen3.6-35B-A3B

Table 5: Controlled low-concurrency deployment replay. Latencies are seconds; Decode is the arithmetic mean of client-observed completion tokens/s over model calls. It is length-sensitive and is not an intrinsic cross-checkpoint kernel-speed measure. TTFT is the first call’s client-observed P95 and includes network and queueing. Each row contains 100 measured Agent requests after 10 warm-up cases. API rows are operational references, not same-hardware model comparisons.

Figure 7: Single-run standalone MTP-loss trajectories for a task-trained policy with a transplanted and subsequently post-trained NextN head, and for the base model with its native factory-trained head. Faint lines show step-level loss; bold lines show a 25-step trailing mean. The runs have different lengths and provide adaptation diagnostics only: they do not compare convergence speed, training stability, sample efficiency, or compute efficiency.

Serving mode	Accuracy	Effectiveness	AVG
MTP Off	95.8	93.8	94.8
MTP On	96.2	94.2	95.2

Table 6: Same-checkpoint

\mathcal{T}_{1}

quality check with MTP Off and On (

n=978

). Both rows use the same Final Evaluation Judge. AVG is the arithmetic mean of Accuracy and Effectiveness.

Checkpoint	$C$	Decode TPS	Speedup	Wall P95 (s)	TTFT P95 (s)	$\leq$ 15s
Qwen3.6-35B-A3B	1	195.42 $\rightarrow$ 215.95	1.11 $\times$	10.172 $\rightarrow$ 9.805	0.508 $\rightarrow$ 0.502	100 $\rightarrow$ 99%
	2	138.81 $\rightarrow$ 165.74	1.19 $\times$	11.504 $\rightarrow$ 10.251	0.623 $\rightarrow$ 0.698	97 $\rightarrow$ 97%
	4	105.04 $\rightarrow$ 113.25	1.08 $\times$	13.408 $\rightarrow$ 16.266	0.643 $\rightarrow$ 1.088	97.3 $\rightarrow$ 92.3%
	8	70.28 $\rightarrow$ 70.19	1.00 $\times$	17.848 $\rightarrow$ 26.136	0.953 $\rightarrow$ 1.519	90 $\rightarrow$ 79%
HAT (Ours)	1	160.30 $\rightarrow$ 271.40	1.69 $\times$	8.98 $\rightarrow$ 8.114	0.499 $\rightarrow$ 0.553	100 $\rightarrow$ 100%
	2	129.22 $\rightarrow$ 196.15	1.52 $\times$	10.074 $\rightarrow$ 9.047	0.692 $\rightarrow$ 0.754	100 $\rightarrow$ 100%

Table 7: MTP trade-off across concurrency. Arrows show Off

\rightarrow

On. Decode TPS is client-observed mean per call; Wall and TTFT are P95 seconds. C=1, 2, and 8 are single runs. C=4 values are means of three run-level statistics, not pooled percentiles.

Preference	Count	Share
Harness better	35	35.0%
Tie	64	64.0%
ReAct better	1	1.0%

Table 8: Human blind-test preference distribution on 100 paired real live-streaming requests.

Attribution category	Count	Share
More accurate input understanding	12	34.3%
More reliable output	8	22.9%
More appropriate scenario behavior	8	22.9%
Higher response quality	5	14.3%
More appropriate tool use	2	5.7%

Table 9: Attribution of the 35 Harness-preferred examples. Categories are assigned from annotator rationales after the blind decision.

Metric per participating UV	Relative uplift	Platform assessment
Item Page View (IPV)	+0.9107%	Significantly positive

Table 10: UV-normalized item-page-view result from the online A/B test in Taobao Live’s production digital-avatar business. The relative uplift compares the Harness treatment with the ReAct control; the significance assessment is reproduced from the experiment platform.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Responsibility	Interfaces	Role
Product retrieval	search_product_by_keyword , get_current_product_info , get_product_info_by_link_id , get_product_extra_info_by_link_id_and_keywords , search_preset_faq	Find the active or referenced product and retrieve catalog, detail-page, knowledge-base, or FAQ evidence.
Pricing and promotion	get_price_info_by_link_id , get_promotion_info	Retrieve SKU-level prices, entitlements, coupons, and room-level promotions.
Action and flow control	change_explain_order , load_skill , get_current_time , refuse_to_reply	Change presentation order, load behavioral instructions, resolve time-sensitive rules, or silently discard meaningless input.
MCP lookup	query_promotion , query_buyer_resource , query_fund_asset , query_coupon_detail	Retrieve campaign conditions and buyer-side resources from the external marketing service.
MCP calculation	calculate_optimal_promotion , calculate_promotion	Compute eligible promotional combinations and resulting prices.

Appendix

Table 11: Working inventory of built-in and MCP-discovered runtime interfaces.

Reply Skill	Response responsibility	Strategy Skill	Tool-chain responsibility
item_qa	Product attributes, specifications, comparison, price, and recommendation.	tool_strategy_benefit	Entitlement, price, and promotion retrieval.
chillchat	Non-product conversation and engagement.	tool_strategy_compare	Multi-product comparison and retrieval.
aftersale	Returns, exchanges, complaints, and other after-sales requests.	tool_strategy_multi	Aggregation and orchestration for multiple comments.
clarification	Follow-up questions when available context is insufficient.	tool_strategy_switch	Presentation-order adjustment.
faq_reply	Responses grounded in configured FAQ entries.	general_discount	Room-level promotion retrieval and response composition.
refusal	Declines for inappropriate or unsupported requests.

Appendix

Table 12: Working Skill inventory and routing responsibilities.

Stage	Harness changes	Acc.	Eff.	Engineering diagnosis
ReAct	Conventional reasoning–action loop without the modular Harness.	80.33	84.58	System baseline.
Harness base	Initial manually maintained Skills and modular runtime.	82.40	87.16	Establishes the editable Harness.
Evolution 1	Add refusal tool and stop-loop Hook; rewrite refusal, chat, and after-sales Skills; add four global constraints.	92.13	84.16	Accuracy rises sharply; internal diagnosis attributes Effectiveness loss to over-triggered refusal.
Evolution 2	Add seven whitelist exclusions before refusal; map explanation triggers and require the relevant tool call.	92.55	92.75	Selected ; restores Effectiveness while retaining Accuracy.
Evolution 3	Add attribution, factuality, parameter-completeness, tool-use, and system-message rules; expand refusal exclusions.	91.51	90.89	Long-tail rules interact and regress both metrics.
Evolution 4	Relax length, tool-trigger, parameter, transaction-intent, and attribution restrictions.	91.51	89.96	Partial rollback does not recover Effectiveness.

Appendix

Table 13: Fixed-policy Harness Evolution record. Values come from six direct evaluation summaries on the same 482-item dev-set. “Selected” is an engineering early-stop decision.

Figure 8: Qualitative prompt-edit cases under the same Harness revision. The edit prohibits disclosure-process language when the available evidence does not support a definite conclusion. In both examples, the Fixed-Harness SFT checkpoint continues to expose internal uncertainty after the edit, whereas HAT follows the revised instruction and gives a direct response or routes the viewer to product details or customer service. Red text marks the disclosure phrases targeted by the edit. These examples illustrate compliance behavior and are not additional quantitative evidence.

Figure 9: Judge alignment to a human-labeled calibration cohort (

n=482

). Accuracy and Effectiveness agreement improve through evidence-tool, rubric, and voting revisions; the cohort evaluates the scoring instrument rather than the policy.

Contrast / dimension	$\Delta$ [95% CI]	$P_{\mathrm{boot}}(\Delta\leq 0)$
IFEval prompt-level accuracy
Fixed-Harness SFT $-$ Base	$-6.47\;[-10.72,\;-2.22]$	$0.9992$
Ours $-$ Base	$+2.03\;[-1.11,\;+5.18]$	$0.1134$
Ours $-$ Fixed-Harness SFT	$+8.50\;[+4.44,\;+12.38]$	$0.0000$
Prompt Robustness
Ours $-$ Fixed-Harness SFT	$+8.46\;[+5.27,\;+11.66]$	$<0.0001$

Appendix

Table 14: Independent paired-bootstrap stability check. Differences are reported in percentage points; intervals are 95% percentile bootstrap intervals over evaluation items.

Route / checkpoint	MTP	Calls/req	Input/call	Output/call
DeepSeek-V4-flash API	–	4.04	6,052	146
DeepSeek-V4-pro API	–	3.67	5,537	119
Qwen3.6-35B-A3B	Off	3.23	5,622	115
Qwen3.6-35B-A3B	On	3.08	5,571	109
HAT (Ours)	Off	3.04	5,679	106
HAT (Ours)	On	3.12	5,786	105

Appendix

Table 15: C=1 workload shape. Values are means over successful complete-Agent requests or their constituent model calls.

Setting	Engine argument	Value
Speculative algorithm	--speculative-algo	NEXTN
Speculative steps	--speculative-num-steps	3
EAGLE top- $k$	--speculative-eagle-topk	1
Draft tokens	--speculative-num-draft-tokens	4
Single-token acceptance threshold	--speculative-accept-threshold-single	0.5
Accumulated acceptance threshold	--speculative-accept-threshold-acc	0.7

Appendix

Table 16: MTP serving configuration used for the reported quality check and deployment measurements.

Checkpoint	Accept len	Accept rate
Qwen3.6-35B-A3B	2.94/2.92/3.58	.735/.730/.890
HAT (Ours)	3.12/3.12/3.62	.780/.780/.910

Appendix

Table 17: SGLang log-sample diagnostics for MTP-On serving. Triples are mean/P50/P95.

Explore similar work

Adaptive Auto-Harness: Sustained Self-Improvement for Agentic System Deployment on Open-Ended Task Streams
Jun 1, 2026Zewen Liu, Zhan Shi, Yisi Sang +7Evolving Harness Multi-Agent Evolution
Emory University · Amazon · The Pennsylvania State University +2
HarnessX: A Composable, Adaptive, and Evolvable Agent Harness Foundry
Jun 12, 2026Tingyang Chen, Shuo Lu, Kang Zhao +11Agent Harness Multi-Agent Evolution
Darwin Agent Team
EvoStreaming: Your Offline Video Model Is a Natively Streaming Assistant
May 11, 2026Zichen Wen, Boxue Yang, Junlong Ke +5Video-Language Models Parameter-Efficient Adaptation
EPIC Lab, Shanghai Jiao Tong University · Tsinghua University · The Hong Kong University of Science and Technology (Guangzhou) +1

Jun 1, 2026, cs.LGJ/K move · Enter open · S save

Adaptive Auto-Harness: Sustained Self-Improvement for Agentic System Deployment on Open-Ended Task Streams

Zewen Liu, Zhan Shi, Yisi Sang, Bing He+6

Emory University · Amazon · The Pennsylvania State University+2

Evolving Harness Multi-Agent Evolution Efficient Task Adaptation Agentic Deployments Recursive Self-Improvement

Open paper AskPDF arXiv

Auto-harness systems such as A-Evolve, GEPA, and Meta-Harness improve LLM agents by optimizing prompts, skills, tools, memories, and supporting infrastructure from execution feedback, but they are typically evaluated on fixed offline benchmarks. Real deployments instead present open-ended task streams: histories grow without a fixed endpoint, heterogeneous tasks require different harnesses, and problem distributions shift over time. These challenges make a single repeatedly and densely updated harness brittle, causing performance degradation as accuracy peaks early and then declines. This motivates sustained harness construction with task-wise adaptation. We introduce Adaptive Auto-Harness, a framework and system for such streams. The framework decomposes the gap to an oracle harness into evolution loss and adaptation loss. The system addresses these losses with a stateful multi-agent evolver, a harness tree with solve-time routing, and human-steering hooks for cases where history lacks the needed signal. Across prediction-market, security-competition, and event-forecasting streams, Adaptive Auto-Harness outperforms five existing auto-harness baselines and ablations attribute gains to better construction, routing, or targeted human steering. Code is available in Link.

Abstract

Explore similar work

Adaptive Auto-Harness: Sustained Self-Improvement for Agentic System Deployment on Open-Ended Task Streams

HarnessX: A Composable, Adaptive, and Evolvable Agent Harness Foundry

EvoStreaming: Your Offline Video Model Is a Natively Streaming Assistant