Mini-SWE-Agent is a minimalist open-source software engineering agent written in Python. In Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents — A Source-Code Study of Eleven Systems, which examines its source at version 2.4.5, it is the “100-line research floor” of an eleven-system corpus: the paper describes it as implementing all seven subsystems it identifies in an Agent Harness in roughly 100 lines — a while loop, one template, one tool, a message list, two limits, no orchestration, and structural typing as its entire extension story. The paper lists it as coming from Princeton and Stanford.
The paper frames the minimalism as deliberate: the design isolates model capability from scaffolding complexity to see how much of one can substitute for the other. It is used throughout that study as the minimal form against which the other systems’ implementations of each subsystem are compared.
Capabilities
The loop is a linear while True that queries the model and executes the returned actions, with no state
machine, no event sourcing and no concurrency. The message list grows without bound and relies entirely on
the model’s native context window, which the paper says makes trajectories perfectly reproducible and easy
to analyse. It has one tool, bash, run as a subprocess call, and its prompt asks for exactly one bash
command per turn, so directory and environment changes have to be prefixed inline because each call runs
in a new subshell. Prompts are Jinja2 templates embedded in YAML configuration and rendered once at
session start, and model access goes through LiteLLM.
Its safety mechanisms are resource limits — a step limit and a cost limit, joined in the 2.4 line by a wall-clock limit and a cap on consecutive malformed responses — plus an opt-in interactive confirm mode. The paper calls this sufficient for benchmark evaluation but inadequate for production use, and notes that even this floor has acquired a minimal stuck detector. Execution can run in pluggable environments including Docker, Singularity and Bubblewrap. Extension works through Python protocols: any class implementing the Model, Agent or Environment protocol can be substituted without inheritance, which the paper calls the lightest-weight extensibility mechanism in its corpus.
Adoption & Ecosystem
The paper uses Mini-SWE-Agent as the main evidence for its observation that loop sophistication does not predict benchmark performance, citing results the project reports on SWE-bench Verified that fall in the same range as far larger systems. It is explicit that such figures are self-reported, obtained on different models and dates, and not a head-to-head comparison. The same study cites it as the model for its own first design recommendation — start with a linear loop — and its 90-line minimum-viable-harness scaffold names Mini-SWE-Agent as the source of the linear loop it combines with other systems’ patterns. For the separate SWE-agent project, see SWE-Agent.