AgentSTAR (also written Agentic STAR), short for Agentic Shape Tracking and Reconstruction from Monocular Videos, is a harness from Amazon FAR (Frontier AI and Robotics) that turns Claude Code or Codex into the operator for articulated 3D object reconstruction from monocular video. Given RGB frames, per-frame object masks and known per-frame cameras, the agent writes and iteratively edits a Blender Python scene.py that builds the object from primitives in a canonical rest pose, poses it per frame with one shared scale, and declares articulated joints and per-frame joint states, while the harness alone owns cameras, lighting, render settings, forward kinematics and export. Built-in sweep/apply pose-search and temporal-report tools drive candidate poses and motion review, and a windows workflow can parallelize per-frame refinement across subagents once geometry is frozen. Each iteration renders and scores the scene against the source frames, gated by a hard 0.98 per-frame silhouette-IoU threshold, turntable coherence, and temporal-adjudication checks, up to a default 300-iteration budget. Output is a watertight GLB with named parts and a pose.json holding the shared scale, per-frame poses and joint states. The project page shows qualitative reconstructions on the ARCTIC and HOT3D hand-object datasets alongside ground-truth hands; the page does not list numeric benchmark scores. It was designed and tested against GPT-5.6-Sol and Claude Fable 5, with the paper's own results using its optional mechanism module enabled (off by default). Setup needs a Linux x86_64 host, an NVIDIA GPU, and a Bubblewrap sandbox that restricts network egress to the model provider's API host; install.sh provisions Blender 4.2.5 LTS and a pi3x camera-estimation environment built on yyfz/Pi3, and the coding agent needs its own supplied API key. Code is MIT licensed; two vendored routines instead carry SciPy's BSD-3-Clause license.