- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
HexAGenT schedules dynamic agent workflows on heterogeneous prefill–decode clusters. It jointly determines call priorities and worker assignments using projected completion times normalized by each workflow’s revealed critical-path length. The planner accounts for queueing, KV-cache reuse, transfer costs, and memory constraints.
Critical strengths
The joint planning of workflow priorities and P-D placement is the paper’s strongest contribution. Incorporating placement costs into workflow prioritization provides a useful way to account for dependencies and hardware heterogeneity within the same scheduling decision.
Critical weaknesses
-
The evaluation does not isolate the benefit of joint planning: the baselines differ in both ordering and placement, and component ablations are missing. Comparisons using the same placement mechanism across ordering policies would make the attribution clearer. An explicitly adapted Autellix priority rule could provide an additional workflow-aware baseline.
-
The refinement criterion needs stronger justification. Eq. (1) targets a percentile of , while refinement prioritizes reducing the maximum projected ratio based on . These criteria can favor different schedules, but the paper does not examine the implications of this trade-off for Req95/Req99.
Questions for authors
-
For ready calls within the same workflow, Eq. (7) orders them by . How does this ordering affect workflow completion in a branching example, and why is refinement disabled for fan-out cases?
-
How are deadlines constructed for SGLang-LLF and Workflow-LLF? Please specify any SLO multiplier and whether the H100 reference is used only for evaluation or also to set scheduling deadlines.
6 comments
Homo results for Qwen. The table reports all methods across the four traces.
Method ShareGPT BFCL LATS Terminal-Bench Workflow-FCFS 4.9928 / 7.8565 5.8287 / 9.3230 1.5303 / 1.5835 2.0317 / 2.7232 Workflow-LLF 3.5562 / 4.1008 5.7637 / 6.8450 1.5303 / 1.5835 1.7095 / 2.0354 Po2C+Call-LLF 4.7672 / 5.8734 5.3650 / 9.0987 1.5155 / 1.5476 1.7502 / 1.8729 ATLAS 2.8335 / 3.3593 5.4773 / 5.7372 1.5089 / 1.5195 1.7381 / 1.9250 AMPD 4.3042 / 6.3095 5.9795 / 8.2259 1.8571 / 2.0294 1.9426 / 1.9759 SAGA 4.7207 / 5.3188 5.9306 / 10.2079 1.5123 / 1.5263 1.5479 / 1.5914 Parrot 3.8260 / 5.6527 5.6934 / 8.7427 1.5260 / 1.5306 1.5318 / 1.5725 Continuum 4.8972 / 7.2803 5.4239 / 7.2604 2.0396 / 2.2319 2.1792 / 2.3852 HexAGenT 2.4637 / 2.5661 3.8878 / 3.9339 1.5025 / 1.5202 1.5660 / 1.5773
Hetero-1 results for Llama. The table reports all nine methods across the four traces.
Method ShareGPT BFCL LATS Terminal-Bench Workflow-FCFS 2.2156 / 3.5789 2.0413 / 2.2433 1.3242 / 1.3508 1.7430 / 2.0680 Workflow-LLF 2.0125 / 2.4009 1.9796 / 2.1133 1.3418 / 1.3804 1.5423 / 1.5873 Po2C+Call-LLF 1.6965 / 2.0295 1.9954 / 2.2860 1.2903 / 1.3372 1.4024 / 1.4605 ATLAS 1.7340 / 1.9729 1.7893 / 1.9759 1.4635 / 1.4805 1.5057 / 1.5470 AMPD 2.0550 / 2.4501 1.8173 / 2.2089 1.5130 / 1.5195 1.8938 / 2.7287 SAGA 1.7831 / 2.4607 1.9357 / 2.2541 1.5003 / 1.5147 1.4605 / 1.4772 Parrot 1.8800 / 2.2791 2.1240 / 2.3906 1.5032 / 1.5198 1.4209 / 1.4477 Continuum 1.8985 / 2.4962 1.8947 / 2.3425 1.5286 / 1.5481 1.4897 / 1.4904 HexAGenT 1.6275 / 1.8163 1.8125 / 1.8547 1.0004 / 1.0063 1.1841 / 1.1894
Hetero-1 results for Qwen. The table reports all methods across the four traces.
Method ShareGPT BFCL LATS Terminal-Bench Workflow-FCFS 1.8886 / 2.5104 20.3224 / 34.2807 1.4585 / 1.4754 1.3535 / 1.3907 Workflow-LLF 1.6479 / 2.2039 19.4959 / 30.1516 1.4585 / 1.4753 1.3195 / 1.3398 Po2C+Call-LLF 1.5064 / 1.5951 20.2461 / 34.0081 1.4337 / 1.4879 1.2827 / 1.3430 ATLAS 1.5173 / 1.5979 14.9956 / 18.8806 1.4621 / 1.4820 1.4367 / 1.4450 AMPD 1.7598 / 1.8058 19.1866 / 31.5433 1.5036 / 1.5219 1.5086 / 1.5297 SAGA 1.5501 / 1.6479 20.3016 / 33.1790 1.4904 / 1.5177 1.3665 / 1.4379 Parrot 1.5152 / 1.5216 19.1765 / 32.1217 1.5254 / 1.5268 1.3631 / 1.4011 Continuum 1.6186 / 2.2644 20.2948 / 33.7471 1.5272 / 1.5797 1.4356 / 1.4363 HexAGenT 1.4139 / 1.5176 11.6166 / 11.6584 0.9977 / 1.0051 1.0314 / 1.0586
Hetero-2 results for Llama. The table reports all methods across the four traces.
Method ShareGPT BFCL LATS Terminal-Bench Workflow-FCFS 26.0393 / 37.3016 0.9839 / 1.2368 1.4121 / 1.4658 5.3190 / 5.9227 Workflow-LLF 11.9316 / 18.6261 0.9839 / 1.2369 1.4307 / 1.5172 2.1625 / 2.7362 Po2C+Call-LLF 25.9613 / 37.3016 0.9672 / 1.2235 1.2656 / 1.3187 1.5862 / 1.8934 ATLAS 15.9441 / 26.3901 1.0554 / 1.1440 1.4585 / 1.4792 1.9331 / 2.1068 AMPD 27.6119 / 49.9255 1.0735 / 1.2261 1.5144 / 1.5337 1.7943 / 1.8331 SAGA 27.7117 / 50.1533 0.9961 / 1.1017 1.4915 / 1.4953 1.7535 / 1.8099 Parrot 26.0738 / 37.3486 1.1009 / 1.2274 1.4875 / 1.4959 1.7232 / 1.7691 Continuum 26.5175 / 36.6627 1.0572 / 1.2018 1.6064 / 1.6497 2.9777 / 4.2461 HexAGenT 8.6013 / 8.6176 0.7104 / 0.7764 0.9852 / 1.0000 1.2361 / 1.2710
Hetero-2 results for Qwen. The table reports all methods across the four traces.
Method ShareGPT BFCL LATS Terminal-Bench Workflow-FCFS 26.2869 / 38.3089 0.9314 / 0.9545 1.4849 / 1.6202 2.6888 / 3.1597 Workflow-LLF 14.7521 / 16.9611 0.9314 / 0.9545 1.4723 / 1.4851 1.4292 / 1.7938 Po2C+Call-LLF 26.7000 / 37.9727 0.8989 / 0.9373 1.4564 / 1.4847 1.1504 / 1.3040 ATLAS 16.9817 / 18.6686 1.0089 / 1.0159 1.4730 / 1.4771 1.4282 / 1.5051 AMPD 28.8821 / 38.5283 1.0029 / 1.0100 1.4893 / 1.5054 1.4775 / 1.6615 SAGA 24.7541 / 34.0269 1.0041 / 1.0082 1.4889 / 1.4937 1.2302 / 1.2970 Parrot 29.6037 / 40.2841 1.0101 / 1.0192 1.4891 / 1.4949 1.2005 / 1.2295 Continuum 26.8737 / 38.3008 1.0069 / 1.0121 1.8650 / 2.0340 1.5684 / 1.9439 HexAGenT 7.8969 / 8.1537 0.6492 / 0.6634 0.8605 / 0.9667 1.0098 / 1.0135
Weakness 2 .
The refinement criterion needs stronger justification. Eq. (1) targets a percentile of , while refinement prioritizes reducing the maximum projected ratio based on . These criteria can favor different schedules, but the paper does not examine the implications of this trade-off for Req95/Req99.
We appreciate the reviewer for highlighting this ambiguity. We would like to clarify that these criteria should favor the same schedules, while refinement should follow the attainment target in Equation (1). The maximum projected ratio in the original writing corresponds to the special case when , but can be other values as well. We apologize for omitting this qualification, which made the criteria appear inconsistent.
Refinement criterion and its relationship to Equation (1). For each candidate schedule, refinement evaluates the -quantile of projected workflow ratios over the same active-workflow set. Choosing or therefore targets the projected 95th or 99th percentile, respectively. These projections use the currently revealed workflow horizon , while final evaluation uses completed workflow durations and the fixed standalone reference $S_w$. The projected percentile thus guides refinement toward the same attainment level specified in Equation (1). Section 5.3 and Algorithm 1 now clarify this parameterization.
We appreciate this suggestion. Section 6.3 and Figure 5 compare HexAGenT with Workflow-LLF using the same joint worker-pair selector, queue/capacity estimates, and runtime reuse. The experiment uses Llama and the Hetero-2 GPU composition.
Ablation configurations. Refinement is held off throughout this existing four-component comparison. Each configuration changes one mechanism:
What these comparisons isolate. The priority comparison isolates ordering; the remaining configurations isolate placement and its information inputs. All configurations use the same workload and evaluation references.
Additional comparison with Autellix. We have also included ATLAS, the program-aware scheduler in Agentix (formerly Autellix), in all 24 main-evaluation scenarios (Section 6.1; Figures 3–4). Its attained-service-based prioritization provides an additional workflow-aware baseline. This comparison evaluates the complete scheduling policies, while the matched-placement comparison with Workflow-LLF in Section 6.3 isolates the contribution of HexAGenT’s priority mechanism. The following tables report all nine methods across the 24 scenarios. Each cell gives α0.95 / α0.99; lower is better.