Seshat Agent

Seshat Agent

Fixes software bugs with a self-improving coding agent that learns reusable reasoning patterns from its own successful runs, for developers and researchers working on repository-level bug fixing.

Made by PCL-RSI-lab 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: Seshat is a coding agent for software engineering that improves itself over time. It runs on MetaEvo, which keeps reasoning patterns taken from past successful attempts and feeds the relevant ones back to the model when a new task arrives. The authors report big gains on two bug-fixing benchmarks when it is added to existing base agents.

Who it suits: It suits developers and researchers who work on automated bug fixing and want to study or try a self-improving agent.

Title: Who uses this model
Industry | Who uses it | How they use it | Example
Software | Engineering teams | Try an agent that fixes bugs in a repository | Point it at a failing issue and review its patch
AI research | Agent researchers | Study self-improvement loops for coding agents | Compare a base agent with and without MetaEvo
Benchmarking | Evaluation teams | Check results against the published evidence | Read the per-instance patches the authors released

If you want a plain chat model or a ready-to-run tool with documented commands, you can skip it.

Where to use it:

• Python on Windows, macOS or Linux, for downloading the files
• Agent setups for software engineering tasks, built on a base model

Title: Jobs it does
Use case | Who | What they get back
Bug fixing | Developers | A proposed patch for a repository issue
Self-improvement research | Researchers | A loop that stores and reuses reasoning patterns
Benchmark runs | Evaluators | Resolve rates on bug-fixing test sets

How to set it up: You need Python and the files from Hugging Face. The visible part of the card gives no run steps, so treat it as research material and check the rest of the card before building on it.

• Python with pip
• The huggingface_hub package
• A base agent model, since the reported results build on Kimi K3 and GLM-5.3

1. Install huggingface_hub.
2. Run the download command to save the files on your computer.
3. Read the full model card for the run steps.

pip install -U huggingface_hub
hf download PCL-RSI-lab/Seshat-agent

How to use it: The card shows no commands in its visible part. From its description, a task such as a bug report goes in, a router picks the most relevant stored reasoning patterns, and the base model reasons with them while a critic watches. Successful runs then feed the knowledge base.

Example: You give the agent a bug report from a repository. It builds a tailored reasoning plan from stored patterns, works on a fix, and a successful fix is added to its knowledge base for later tasks.

For newcomers and experts: Think of it as a coder that keeps notes on what worked and reuses those notes on new problems. A first small step is reading the card's results table and the released evaluation evidence.

MetaEvo stores compact reasoning primitives, such as decomposition, self-verification and tool-calling patterns, rather than raw text memory. At run time it injects a selected path into the model layer by layer, and a meta-critic can continue, re-route or re-plan. Offline, it mines new primitives from successful traces.

Problem comes in
Router picks stored reasoning patterns
Plan is injected into the model
Meta-critic watches each step
Successful runs update the knowledge base

Good to know: The authors report that Seshat raises a Kimi K3 agent from 57.7% to 94.94% on SWE-Bench Pro, and a GLM-5.3 agent from 59.3% to 90.3% on DeepSWE-v1.1. They publish patches and per-instance results as evidence. These are their own results.

Title: Resolve rate reported by the authors
Kimi K3 alone on SWE-Bench Pro: 57.7%
Kimi K3 with Seshat on SWE-Bench Pro: 94.94%
GLM-5.3 alone on DeepSWE-v1.1: 59.3%
GLM-5.3 with Seshat on DeepSWE-v1.1: 90.3%

• License: Apache 2.0.
• Cost to run: none named for Seshat itself, though the results used Kimi K3 and GLM-5.3 as base agents.
• Data it reaches: the files come from Hugging Face, and the agent works on the software project you give it.
• Not for: anything the card rules out, since its visible part names no limits.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Downloads in the last 30 days: 0
• Likes: 7
• License: apache-2.0
• Tags: transformers, agent, recursive-self-improvement, self-evolving, metaevo, meta-knowledge, swe-bench, swe-bench-pro, deepswe, code-agent, kimi, text-generation

Links

• Model: https://huggingface.co/PCL-RSI-lab/Seshat-agent
• Files: https://huggingface.co/PCL-RSI-lab/Seshat-agent/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download PCL-RSI-lab/Seshat-agent

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