From task execution to human-centric outcomes: relationship based context management,
trust and intent modeling, and continuous decision progression —
turning every interaction into real business results.
Empathy Model +
Vertical Agentic Model + Agent Harness
→ HUMAN-CENTRIC LONG HORIZON AGENT
01 / Beyond Task Delivery
把任务做成，把客户旅程持续推进Deliver the task. Keep the customer journey moving.
一次方案交付之后，客户可能还在犹豫，审批可能仍在等待。HUMAN-CENTRIC LONG HORIZON AGENT 将可靠的任务执行连接到客户的长期决策：理解客户变化、延续任务目标，让任务成果转化为持续的业务进展。
After an offer is delivered, a customer may still be deciding or an approval may still be pending. HUMAN-CENTRIC LONG HORIZON AGENT connects reliable task execution to long-term customer decisions through evolving customer understanding, lasting relationships, and continued business progress.
能力维度Dimension
通用任务 AgentGeneral-purpose task agent侧重复杂任务的执行与交付Focused on complex task execution and delivery
HUMAN-CENTRIC LONG HORIZON AGENT在可靠交付之上，持续推进客户与业务Build on reliable delivery to advance customers and business
任务目标Task objectives
通用任务 AgentGeneral-purpose task agent完成明确的任务目标Complete a defined task objective
围绕明确的任务要求规划与执行，以交付约定结果为目标，解决客户提出的具体问题。
Plan and execute against defined requirements to deliver the agreed result and resolve a specific customer request.
HUMAN-CENTRIC LONG HORIZON AGENT跨时间接续执行，持续推进长期目标Resume execution across time to advance long-term goals.
Organize tasks around long-term goals and preserve context and progress across interactions. Resume after waits, external feedback, or changed conditions to sustain progress over time.
Combine domain knowledge, business rules, and tools to understand task constraints and conditions for action. Choose next steps based on user needs, task state, and execution feedback, adapting decisions as conditions change.
共情理解Empathetic understanding
通用任务 AgentGeneral-purpose task agent理解任务需要什么Understand what the task requires
从指令与对话中识别需求，为当前任务提供上下文。
Identify needs in instructions and conversation to inform the current task.
HUMAN-CENTRIC LONG HORIZON AGENT记住一个人，理解每一次变化Remember the person. Understand what changes.
Bring an empathy model into customer understanding: remember preferences and commitments across conversations, recognize changes in intent, budget, trust, and concerns, and adapt offers and communication to the customer's situation.
价值衡量Value measures
通用任务 AgentGeneral-purpose task agent关注任务是否完成Measure whether the task was completed
主要衡量完成率、结果质量、执行效率与成本，判断一次任务的交付表现。
Assess task delivery through completion rates, result quality, execution efficiency, and cost.
HUMAN-CENTRIC LONG HORIZON AGENT自动持续优化，提升任务效果Continuously optimize automatically to improve task outcomes.
Evaluate delivery quality, user needs, and progress toward long-term goals, using execution feedback to guide improvement. Diagnose failures, refine candidates, and run regression evaluations to improve strategies, skills, and future task performance.
02 / Open-source Feedback Loop
从模型到业务，再以反馈持续改进From models to business, improved through feedback
以垂域模型为基础，连接长程任务、语音交互与评测自进化；未来将进一步探索训练回流，持续优化模型能力。
Connect domain models with long-horizon tasks, voice interaction, and evaluation-driven evolution. Future work will explore feeding these insights back into model training.
Organize context around memory, state, beliefs, goals, and policy. Connect human understanding, action, human oversight, and outcome feedback to advance goals across conversations.
跨会话记忆：延续客户关系Cross-session memory: sustain the relationship
Retain past exchanges, preferences, and commitments, then combine them with the current conversation. Pick up where the relationship left off, so customers do not need to explain everything again.
客户状态建模：理解变化中的人Human state modeling: understand change
Applications can use context extensions to organize intent, budget, trust, urgency, and buying stage. Separate historical facts, current state, and working beliefs to inform the next action.
长程任务推进：按策略持续跟进Long-horizon progression: follow up over time
Persist task state across days. Break goals into stages and use scheduling and wakeup policies to continue after new documents, approval feedback, or a new offer, closing the loop from observation to action and feedback.
Connect steps such as checking eligibility, calculating offers, and arranging follow-up. Tools execute actions in business systems; suspend for asynchronous approvals or external feedback and resume when results arrive.
实时交互：响应新的意图Live interaction: respond to new intent
Users can interrupt or add information while the agent runs. New messages join the active run, allowing the agent to adjust its next step as needs and constraints change.
人机协同：关键时刻交给人Human-in-the-loop: involve people when it matters
通过工具挂起与恢复机制，支持应用在合规风险、重大权益或例外场景引入人工确认；携带必要上下文完成接力，处理后由 AI 继续推进。具体介入规则由业务应用定义。
Tool suspension and resumption let applications involve people for compliance risks, material rights, or exceptions. Carry the necessary context into the handoff and let AI continue afterwards. Applications define intervention policies.
快速接入：配置业务知识与行动Fast onboarding: configure knowledge and actions
Compose models, tools, skills, and knowledge bases in YAML to connect domain expertise and business actions. The configuration platform supports online editing and dry-run validation; save and restart to apply.
评测驱动进化：从反馈中改进Evaluation-driven evolution: learn from feedback
Full traces support evaluation and failure diagnosis; successful patterns become reusable skills. Release candidates after regression evaluation. Applications can define evaluation criteria around appointments, visits, and conversion.
04 / Evolution
评测驱动的安全进化，而不是让 Agent 随意改自己
Evaluation-driven safe evolution, not an agent rewriting itself at will
HUMAN-CENTRIC LONG HORIZON AGENT evolves through a data-operations loop: full production traces are evaluated in
layers and split — successes are distilled into reusable skills, while failures are
diagnosed into challenger candidates. A candidate ships only after beating the current
best and passing golden-set regression. Test is a separate manual action and never
participates in selection.
方法设计服务三个目的：及时预警线上同类型大批量错误；沉淀优秀案例、复用成功经验；
让线上 case 自动修复并产出优化后的新版本。
The methodology serves three goals: early warning when the same kind of failure repeats at
scale in production; distilling successful cases for reuse; and turning online failures into
an improved new version automatically.
Self-Evolution Methodology · 自进化方法设计
运行与评测Execution & evaluation
业务 AGENTBUSINESS AGENT
线上运行
online serving
→
TRACE 记录TRACE LOGGING
完整会话 · 工具/Skill 调用 · 知识库检索 · Memory
full sessions · tool/skill calls · KB lookups · memory
The latest experiments cover the banking scenario in τ³-bench and the office and information
retrieval scenarios in EvoAgentBench. Across six configurations, evolved candidates improve all
40 before-and-after comparisons reported in the source table. Because the two benchmarks use
different tasks and scoring protocols, their absolute scores are presented separately rather
than blended into one ranking.
每个雷达独立阅读：灰色圆点为 Base，蓝色菱形为 Evolved；轴标签给出该榜单内部的平均提升。
形状扩张表示本榜单内的进化效果，不用于两个榜单之间的排名。
Read each radar independently: gray circles are Base and blue diamonds are Evolved; axis labels
show mean uplift within that benchmark. Shape expansion shows within-benchmark evolution, not a cross-benchmark ranking.
各榜单按其纳入配置取算术平均；绿色数值为相对 Base 的百分点提升（pp）。
该表用于查看榜单内部的进化效果，不用于比较两个榜单的绝对难度或排名。
Values are arithmetic means across the configurations included in each benchmark; green values
show percentage-point uplift over Base. Read gains within each row rather than ranking the two benchmarks by absolute score.
Benchmark
统计范围Scope
Base · 原始版本Initial
Evolved · 自进化后After evolution
Train
Validation
Test
Reward
Train
Validation
Test
Reward
τ³-bench
4 组模型 × 检索配置均值Mean of four model × retrieval configurations
0.500
0.397
0.356
0.485
0.735+23.5pp
0.516+12.0pp
0.544+18.8pp
0.588+10.3pp
EvoAgentBench
2 个场景均值：办公 + 信息检索Mean of two scenarios: office + information retrieval
0.560
0.521
0.326
0.463
0.668+10.8pp
0.722+20.1pp
0.452+12.7pp
0.602+14.0pp
Benchmark 01
τ³-bench · Banking：多轮工具调用与长链路执行
τ³-bench · Banking: multi-turn tool use and long execution chains
Derived from
tau2-bench,
τ³-bench uses strict pass/fail scoring to test whether an agent can complete customer requests
through multi-turn conversations and domain tools. This experiment covers two model families and
two retrieval modes; validation stays a black-box evaluation and the test split remains held out.
真实运行证据：tau3_banking_t34v23 任务 12 个候选的线性进化链——训练集与验证集分数随迭代持续爬升，最终候选 0.735，相对初版 +26.4ppEvidence from a real run: the 12-candidate linear evolution chain on tau3_banking_t34v23 — train and validation scores climb iteration over iteration, ending at 0.735, +26.4pp over the initial version图表数据表Chart data table
12 个候选（按时间顺序）的训练集与验证集分数；“—”表示该候选未跑验证集
Train and validation scores of the 12 candidates in time order; “—” means the validation set was not run for that candidate
#
时间Time
Train
Validation
1
09/19 23:09
0.471
0.261
2
09/20 01:04
0.471
0.261
3
09/20 03:45
0.471
0.217
4
09/20 06:27
0.559
0.337
5
09/20 08:28
0.559
0.283
6
09/20 10:43
0.441
—
7
09/20 11:42
0.559
0.413
8
09/20 18:10
0.588
0.424
9
09/20 22:37
0.471
—
10
09/21 00:27
0.588
0.413
11
09/21 04:11
0.647
0.478
12
09/21 11:56
0.735
—
τ³-bench 四组配置：自进化前后全部指标提升
τ³-bench: every metric improved across all four configurations
On τ³-bench · Banking, evolved candidates improve all 16 sampling-consistency comparisons across
Pass^1–Pass^4 (+4.9 to +18.1pp), as well as every train / validation / test and reward comparison;
no regression was observed on the held-out test split in this experiment.
原始版本 → 自进化后，数值保留三位小数。Base → Evolved, rounded to three decimal places.
模型 / 检索方式Model / retrieval
Pass1
Pass2
Pass3
Pass4
Reward
qwen3.8-27Bbm25
0.436→0.598
0.328→0.509
0.271→0.446
0.243→0.402
0.436→0.600
qwen3.8-27Bgolden_retrieval
0.558→0.637
0.497→0.562
0.454→0.513
0.423→0.474
0.560→0.637
deepseek-v4-flash-0713bm25
0.339→0.421
0.255→0.365
0.216→0.342
0.195→0.337
0.349→0.442
deepseek-v4-flash-0713golden_retrieval
0.591→0.663
0.536→0.606
0.503→0.564
0.484→0.533
0.595→0.674
表 2 · 训练集、验证集与测试集Table 2 · Train, validation & test
原始版本 → 自进化后，数值保留三位小数。Base → Evolved, rounded to three decimal places.
模型 / 检索方式Model / retrieval
训练集Train
验证集Validation
测试集Test
qwen3.8-27Bbm25
0.471→0.794
0.348→0.457
0.300→0.550
qwen3.8-27Bgolden_retrieval
0.559→0.794
0.446→0.641
0.350→0.625
deepseek-v4-flash-0713bm25
0.353→0.559
0.217→0.337
0.350→0.400
deepseek-v4-flash-0713golden_retrieval
0.618→0.794
0.576→0.630
0.425→0.600
四组配置为模型 × 检索方式的组合；顶部总览图与汇总表中的 τ³-bench 数据为四组算术平均；pp = 百分点。Each configuration combines a model and a retrieval method. The overview chart and summary table show the arithmetic mean across all four τ³-bench configurations; pp = percentage points.
提升热力图：4 组配置 × Pass^1–4 共 16 项对比全部改善（+4.9 ~ +18.1pp），颜色越深提升越大——没有一项回退Improvement heatmap: all 16 comparisons (4 configurations × Pass^1–4) improved, +4.9 to +18.1pp — darker means larger gain, not a single regressionReward 得分：四组配置从灰点（原始版本）到绿点（自进化后）的跨度一眼可见，提升 +7.7 ~ +16.4ppReward score: the span from each gray dot (base) to its green dot (evolved) is visible at a glance — gains of +7.7 to +16.4pp
多次采样一致性（Pass^1–Pass^4）：四组配置（模型 × 检索方式）逐项对比，每一格蓝柱均高于灰柱，红色数字为提升幅度（pp）Multi-sample consistency (Pass^1–Pass^4): four configurations (model × retrieval) compared item by item — every blue bar beats its gray baseline, red numbers are gains in pp
训练 / 验证 / 测试集与 Reward 得分：提升不限于训练集，held-out 测试集与 Reward 同样提升Train / validation / test and reward score: gains are not confined to the training split — the held-out test split and the reward score improve as well
多次采样一致性曲线：横轴为采样次数 k，蓝线（自进化后）全程位于灰线（原始版本）上方，阴影面积即整体差距，红色数字为各档提升（pp）Multi-sample consistency curves: k on the x-axis; the blue line (evolved) stays above the gray line (base) end to end — the shaded area is the overall gap, red numbers are per-tier gains in pp
On qwen3.8-27B, both scenarios produced an evolved candidate after 11 iterations. The office
scenario starts from strong train / validation baselines (0.781 / 0.792) and still lifts
reward from 0.686 to 0.724 (+3.8pp). In information retrieval, validation rises from 0.250
to 0.611 (+36.1pp) and reward from 0.240 to 0.481 (+24.1pp).
EvoAgentBench：办公与信息检索场景的 Train / Validation / Test / Reward 全部提升EvoAgentBench: train, validation, test and reward all improve in both office and information-retrieval scenarios
两类场景的自进化表现Evolution results across two scenarios
qwen3.8-27B 在两个场景中自进化前 → 后的得分，数值保留三位小数；pp = 百分点
Scores before → after evolution for qwen3.8-27B in two scenarios, rounded to three decimals; pp = percentage points
Evidence & reproduction: τ³-bench uses 34 train and 23 validation cases;
the test split stays held out and never participates in selection. To reduce network jitter
and model randomness, validation runs with num_trials=4 and counts toward selection
only when the valid-trial ratio is ≥ 0.85; genuine non-infrastructure failures are not re-run.
06 / Scenarios
两个真实业务，跑在同一套系统能力上
Two real businesses, one set of runtime capabilities
客服与电销是两类典型的长程交互业务：一个从全渠道服务请求进入，一个跨 14 天持续推进成交。
两个场景中的每个功能模块，都映射到 HUMAN-CENTRIC LONG HORIZON AGENT 的一项运行时能力，而不是一次性定制逻辑。
Customer service and telesales are two typical long-horizon businesses: one starts from
omnichannel service requests, the other advances a deal across fourteen days. Every module
in both scenarios maps onto a HUMAN-CENTRIC LONG HORIZON AGENT runtime capability — no one-off bespoke logic.
场景一 · 客服：AI 知识检索服务中枢
Scenario 1 · Customer service: the AI knowledge hub
八大功能模块覆盖五类客服业务场景：日常咨询 · 技术答疑 · 服务类 · 定期招揽 · 诉求变更Eight modules cover five scenario types: general inquiries · tech Q&A · service requests · periodic outreach · request changes
场景二 · 电销：一通电话的 14 天旅程
Scenario 2 · Telesales: the 14-day journey of one call
AI × 客户 × 销售坐席三方协同的全流程：客户、AI 与销售坐席共享同一目标和上下文；图片输入进入
多模态理解，外部业务系统可以异步回调并恢复同一次运行，复杂节点允许人工接管后由 AI 带着上下文
接回，而客户沉默时调度器会在未来主动开启新的运行继续推进。
One call, three parties — AI, customer, and human sales agent — sharing one goal and one context.
Images flow into multimodal understanding; external business systems call back asynchronously and
resume the same run; humans take over at complex moments and the AI re-joins with full context;
when the customer goes silent, the scheduler opens a new run in the future to keep things moving.
一通电销电话的 14 天旅程The 14-day journey of one call
AI 坐席AI agent客户（用户）Customer销售坐席Sales agent
1AI 外呼 · 三方同场AI outreach, three-party liveDAY 0
AI 坐席AI →
客户customer
销售坐席全程旁听 · 随时介入
sales listens in and can step in anytime
系统按外呼策略自动拨打，多轮对话识别意向
Policy-driven auto-dial with multi-turn intent qualification
三方协同three-party collab
→
2客户发来证件照片Customer sends ID photosDAY 0
证件照片ID photos →
关键字段抽取field extraction
身份证 · 行驶证 · 车辆登记证
ID card · driving license · registration
Asset 统一安全存储，多模态引擎识别图片并抽取字段
Assets stored safely; the multimodal model extracts key fields
多模态理解multimodal
→
3高意向 · 人工接管High intent · human takeoverDAY 1
AI →
销售坐席sales →
客户customer
对话摘要随接管同步
conversation summary synced on handoff
沟通结束后 AI 携带完整上下文接回，继续执行
Afterwards the AI rejoins with full context and continues
人机切换human handoff
↓任务持续推进 · 状态跨天保持the journey advances · state persists across days
Task state persists across days, advancing stage by stage to closure
长周期任务long-horizon task
一张图看懂系统核心能力：从 Day 0 首呼到 Day 14 成交回访、闭环归档The core capabilities in one journey: from the Day-0 cold call to the Day-14 follow-up and closed loop
07 / Quickstart
从运行到进化，按模块开始Start by module — from runtime to evolution
先完成一次公共准备，再按目标选择模块。启动后端只需要 Python 3.13+、uv
和 Docker；使用 Web 控制台时再准备 Node.js。实际调用模型前，请在 .env
中填写对应的模型配置。
Complete the shared setup once, then start only the modules you need.
The backend requires Python 3.13+, uv, and Docker; the web console additionally requires Node.js.
Add the relevant model configuration to .env before making real model calls.
公共准备Shared setup
安装依赖并创建本地配置Install dependencies and create local config
Choose a task from eval_tasks/, run real agent cases, and inspect rule-based and LLM Judge reports.
EVAL_TASK_ID=your-task-id
curl -X POST \
"http://127.0.0.1:10060/v1/evaluation/tasks/${EVAL_TASK_ID}/runs"
可视化入口：Visual entry point: /platform/evaluation
04 / Evolution
开启 Agent 自进化Start Agent evolution
基于 Train 与 Validation 证据自动诊断、修改、复评并保留更优候选版本。
Use Train and Validation evidence to diagnose, modify, re-evaluate, and retain better candidates.
EVOLUTION_TASK=evolution_tasks/your-task
uv run evolution run "$EVOLUTION_TASK" \
--runtime-config config/runtime.yaml
uv run evolution status "$EVOLUTION_TASK"
