agents

agents

Authoritative registry of agent system profiles, execution harnesses, deliberation search architectures, and reference platforms.

Classes & Registries

Object Type Description
Agents Registry Root registry grouping all agent subclasses (Coding, Deliberation, MultiAgent, Interactive, Platforms).
AgentPlatforms ReferencePlatforms Canonical reference platforms binding compute nodes, models, and sandboxes.
AgentArchitecture BaseModel Typed schema for agent execution harnesses, context boundaries, and sandbox metrics.

Sub-Registries

Agents.Coding (CodingAgents)

  • SWE_Bench_Runner: Reference SWE-bench coding harness hosted on Systems.Nodes.HGX_H100_EPYC with Firecracker microVM sandboxing (5 ms startup, 15 ms snapshot restore, 512 MiB footprint, 30 max steps).
  • SWE_Bench_Workstation: Local unified-memory developer baseline hosted on Systems.Nodes.Workstation_M3Max paired with Models.Language.Llama3_8B.

Agents.Platforms (AgentPlatforms)

  • Cloud_DGX_H100: Enterprise cloud cluster execution envelope (SWE-bench harness on HGX H100 with dual 96-core EPYC CPUs and 8× H100 SXM5).
  • Workstation_Apple: Local developer workstation envelope (SWE-bench harness on Apple M3 Max with 16 CPU cores and 128 GiB unified memory).

Agents.Deliberation (DeliberationAgents)

  • TreeSearch: Monte Carlo Tree Search (MCTS) deliberation agent with 8 branches, 64k context window, and PRM verification.

Agents.MultiAgent (MultiAgents)

  • SupervisorWorker: Hierarchical multi-agent coordinator with 32k context window and coordination overhead parameter \(\beta = 0.030\).

Agents.Interactive (InteractiveAgents)

  • StreamingVoice: Speech-to-speech conversational pipeline with 8k context window and real-time latency budget (TTFT \(\le 200\text{ ms}\)).

Python Usage

import mlsysim

# Access reference platforms
cloud_platform = mlsysim.AgentPlatforms.Cloud_DGX_H100
local_platform = mlsysim.AgentPlatforms.Workstation_Apple

# Inspect sandbox and host parameters
print(cloud_platform.serving_node.host_cpu)         # Dual AMD EPYC 9654
print(cloud_platform.serving_node.host_cpu_cores)   # 192
print(cloud_platform.sandbox_startup_latency)       # 5.0 ms
print(cloud_platform.sandbox_memory_footprint)      # 512 MiB
Back to top