AgentForesight-7B

AgentForesight-7B

Audits a multi-agent run while it unfolds and flags the earliest decisive error, naming the step, the agent and a short reason. For developers who build and debug teams of agents.

Made by ZBox008003 on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: AgentForesight-7B watches the execution trail of a multi-agent system as it grows and decides whether to let it continue or raise an alarm. An alarm names the earliest decisive error, the step it happened at, the agent responsible and a short reason. It is built on Qwen2.5-7B-Instruct and trained with two preference and policy optimization stages.

Who it suits: Developers who run several agents together and want to catch the first thing that went wrong, and researchers who study failures in agent systems.

Industry | Who uses it | How they use it | Example
Software | Agent framework developers | Check each prefix of a run for the first decisive mistake | An auditor that flags step 4 and the agent that caused it
AI quality | Evaluation engineers | Score runs in a batch and read the per-trajectory results | A results file for a set of recorded runs
Research | Multi-agent researchers | Study early failure prediction on the AFTraj-2K dataset | Run the paper split with the macro-domain option

You can skip it if you run a single agent or only need a general chat model, since this one is trained for auditing agent trajectories.

Where to use it: It runs from the command line on Windows, macOS or Linux with Python, through a runner in the maker's GitHub repository.

• A terminal with Python and pip
• The transformers library, with the checkpoint, tokenizer and chat template in one repository
• Your own multi-agent logs, once they follow the input format in the runner's README

Use case | Who | What they get back
Online auditing | Agent developers | SAFE, or the first error step with its agent and reason
Batch evaluation | Evaluation engineers | Per-trajectory predictions and summary metrics in two files
Paper reproduction | Researchers | Results on the paper split

How to set it up: You need Python with pip, git, and enough disk space for the weights, about 15.2 GB in BF16.

• Python with pip and git
• About 15.2 GB for the BF16 weights, or about 3.8 GB at 4-bit as an estimate
• The accelerate package, installed with the requirements

1. Clone the maker's AgentForesight code repository from GitHub and move into its folder.
2. Install the requirements together with accelerate.
3. Download the AFTraj dataset with the one-line Python call.
4. Run the local inference runner. The model weights download by themselves when it starts.

cd AgentForesight
pip install -r requirements.txt accelerate
python -c "from huggingface_hub import snapshot_download; snapshot_download('ZBox008003/AFTraj', repo_type='dataset', local_dir='data/AFTraj')"

How to use it: Run the inference runner. It shows the model each prefix of a trajectory in order. It writes per-trajectory predictions and summary metrics to files, and you drop the one-trajectory limit to process more.

python -m inference.infer_local \
  --model-path ZBox008003/AgentForesight-7B \
  --data-dir data/AFTraj \
  --output-dir outputs/agentforesight \
  --max-trajs 1

Tool | What it does
python -m inference.infer_local | Runs the auditor over trajectories
--max-trajs 1 | Limits the run to one trajectory, remove it to process more
--paper-test-split --macro-domain | Evaluates the paper split
per_sample.jsonl | Holds the per-trajectory predictions
results.json | Holds the summary metrics

Example: Run the command above on one trajectory from the AFTraj data. The model answers SAFE when it sees no decisive error. When it does see one, you get the earliest bad step, which agent caused it and a short reason, and these land in the two output files.

For newcomers and experts: Think of it as a reviewer reading over the shoulder of a team of AI agents. It reads their log one step at a time and either says all is fine or points at the first step that doomed the task. A first small try is the one-trajectory run shown above.

For experts, the checkpoint comes from Qwen2.5-7B-Instruct, trained with boundary-pair preference optimization followed by group relative policy optimization. Auditing is online, so it sees only the prefix up to each step and has to commit to continuing or flagging. That fits live monitoring better than a verdict after the run ends, and the trade is that the model must judge with less context.

Feed the runner a recorded trajectory
The runner shows the model one prefix at a time
The model answers SAFE or flags a step
Flags carry the agent and a short reason
Results go to two files in the output folder

Good to know: The weights, the code and the dataset carry different licenses, and the model checkpoint ships with its tokenizer and chat template.

• Input format: The runner's README holds the input format and more examples, so check it before using your own logs.
• Output: Predictions go to per_sample.jsonl and summary metrics go to results.json under the output folder you pass in.
• Starting small: The command uses one trajectory by default, and you remove the flag for a full run.

• License: Apache 2.0 covers the weights. The runner code uses the MIT license and the AFTraj-2K dataset uses CC BY 4.0.
• Cost to run: none named, the source lists no API key, account or paid service.
• Data it reaches: the model weights and the AFTraj dataset, both downloaded from Hugging Face, plus the files the runner writes to your output folder.
• Not for: none named in the source, which describes an auditor of multi-agent runs and not a general chat model.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 7.6B
• Weights: about 15.2 GB in BF16; about 3.8 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 663
• Likes: 3
• License: apache-2.0
• Languages: en
• Based on: Qwen/Qwen2.5-7B-Instruct
• Quantized versions: 1
• Tags: transformers, safetensors, qwen2, text-generation, multi-agent-systems, online-auditing, failure-attribution, conversational

Links

• Model: https://huggingface.co/ZBox008003/AgentForesight-7B
• Files: https://huggingface.co/ZBox008003/AgentForesight-7B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ZBox008003/AgentForesight-7B

Then follow the usage in its model card. Nothing in this kit runs by itself.