AgentX: Agent-Driven Self-Iteration for Industrial Recommendation¶

简体中文

Reproduction level: core-loop mini-suite. The local controller implements Brainstorm, Developing, Evaluation, and Harness Evolution, and checks how experiment reuse reduces execution cost.

Paper information¶

Field	Details
Paper	arXiv 2606.26859
Organization	Kuaishou (credited to the AgentX Team).
First public release	2026-06-26 (arXiv v1)
Authors' code	No public author implementation was found as of 2026-08-24.
Adapter / method	`agentx`
Local reproduction code	`src/auto_research/agent_research/latest_20260824.py`

Original paper¶

Motivation and method¶

Traditional recommender iterations require engineers to connect hypotheses, production code, live A/B tests, and attribution, and lessons are difficult to accumulate. AgentX creates a four-stage loop. Brainstorm draws evidence-backed candidates from prior experiments, system knowledge, analysis, and papers. Developing implements and verifies them within repository constraints. Evaluation uses guarded online A/B with vetoes. SGPO updates the harness from successful and failed trajectories.

flowchart LR
  Q[Business objective] --> B[Brainstorm: bounded proposals and multiple evidence sources]
  B --> D[Developing: repository-constrained implementation and local checks]
  D --> E[Evaluation: safe traffic and A/B guardrails]
  E --> K[Archive positive and negative experiments]
  K --> H[Harness Evolution / SGPO]
  H --> B

Key figure from the paper¶

Figure 1: The workflow, key stages, and data flow. From the original paper; copyright remains with the authors.

Core equations¶

Proposal scores weigh objective fit, business value, feasibility, handoff completeness, and evidence, then subtract risk:

\[ E(c\mid q)=\sum_{k\in\mathcal K}\alpha_k(q,c)e_k(c),\qquad S(c\mid q)=\lambda_oO+\lambda_bB+\lambda_fF+\lambda_hH+\lambda_eE-\lambda_rR. \]

A significant main-metric gain is insufficient when any critical guardrail vetoes release. Both success and failure traces enter the knowledge base for later proposals and harness updates.

Paper results¶

Three weeks of production deployment covered Kuaishou's main feed and local-life recommendation. Three workers turned 374 ideas into 10 launchable experiments. Concurrency was 8× the manual process, weekly throughput per worker doubled, and business value was 3.7× the human engineer comparison. Launched experiments report +0.561% app time and more than RMB 100 million annualized revenue. This is production-loop evidence, not merely an offline benchmark.

Local reproduction¶

Of 120 deterministic planbench-mini episodes, 12 first-seen task types execute proposal, implementation, double local checks, global guardrails, archival, and policy updates; the remaining 108 reuse archived solutions. Joint success is 1.0000 and average proxy cost 0.5870.

Metric	Local result
Joint success	1.0000
First-time loop / reuse	12 / 108
Bounded proposal explorations	36
Local / global verifier calls	132 / 120
Archival writes / policy updates	12 / 120
Average proxy cost	0.5870

See stable metrics and batch index.

auto-research agent-eval --method agentx --benchmark planbench-mini \
  --episodes 120 --memory-size 24 --seed 42

Limits¶

Without Kuaishou's internal repositories, traffic, users, and SGPO large-model training, this cannot reproduce business uplift. The mini-suite checks loop mechanics, reuse, archived outcomes, and verifier gates. Proxy cost measures control flow, not productivity, revenue, or app time.