Skip to main content

AI & Agent Guide

TensorRT-Model-Connect is designed to be navigable by humans and agents. The machine-readable index is llms.txt. Repository instructions in the repository AGENTS.md remain authoritative; this page is an orientation, not a substitute for reading the instructions that apply to the current checkout and path.

Required agent behavior​

Before changing or running anything, an agent must:

  1. read the applicable AGENTS.md files and obey the closest scoped file;
  2. inspect the current branch, status, remotes, and user-owned changes;
  3. use current model descriptors and E2E manifests instead of guessing model IDs, paths, precision, topology, or support;
  4. separate source/static proof, GPU execution, model parity, performance, and publication evidence in its report; and
  5. state which meaningful validations were not run.

Safety boundaries​

An agent must not:

  • delete, overwrite, reset, rebase, commit, push, open/merge a PR, or publish artifacts unless the user authorized that action and scope;
  • weaken test criteria, tolerances, or gates for the purpose of making CI pass;
  • silently replace an exact checkpoint with a same-family model or unpinned revision;
  • treat a parser option, source file, manifest, or skipped test as target qualification;
  • expose credentials, gated model assets, private URLs, or retained internal artifacts in public output; or
  • mix native runtime support with a platform-specialized provider claim.

When requested work conflicts with one of these boundaries, stop, show the concrete conflict, and ask for human direction.

Agent workflow​

Read instructions
  → inspect exact repository state
  → identify model/task/runtime owner
  → make the smallest scoped change
  → run the smallest meaningful validation
  → report evidence and unrun boundaries
  → wait for authorization before external or destructive actions

Useful source-of-truth commands:

git status --short --branch
python3 tools/model_ci.py validate
python3 tools/check_doc_file_references.py --strict website/docs

AI-native quick start prompt​

Give a capable coding agent this goal from a clean terminal:

/goal Use the current TensorRT-Model-Connect checkout, or clone
https://github.com/NVIDIA/TensorRT-Model-Connect.git if none is provided. Read
AGENTS.md, then follow website/docs/getting-started/source-build.md and
website/docs/getting-started/quick-start.md exactly. Do not modify source,
tests, Dockerfiles, git history, or remote state. Report the selected GPU,
exact commands, bundle path, inference output, and every deviation from the
documentation.

The agent should use Build from Source and Quick Start, not invent a parallel setup.