StackMap
Subscribe

mini-swe-agent vs prime-agent

The 100-line agent from the SWE-bench team: >74% on SWE-bench Verified with no tools but bash, no config sprawl — the reference minimal harness, adopted by Meta, NVIDIA and Ramp. — versus — Self-improving coding/research agent around a Recursive Language Model: persistent IPython as the core tool, programmatic subagents, durable harness state it refines via evidence-backed /refine.

The curated verdict

The two poles of coding-agent design: mini-swe-agent proves 100 lines and bash suffice; Prime Agent bets on a persistent REPL, subagents and self-refining harness state.

mini-swe-agentprime-agent
Stars6.3k6.4k
Forks866515
LanguagePythonTypeScript
LicenseMITMIT
Last activity6 days ago2 days ago
Topicscoding, agentscoding, agents
Curated connections63

mini-swe-agent — the curator's take

The existence proof that most harness complexity is optional: the team that built SWE-bench and SWE-agent asked what a 100x simpler agent loses — the answer is almost nothing (>74% Verified), which is why it became the standard baseline harness for benchmarking models (Ramp's SWE-bench, DeepSWE — where it beats Claude Code and Codex as a harness). Read it to understand agents; use it to evaluate models fairly. NOT a daily driver: no MCP, no skills, no IDE plumbing — by design. If you're choosing a tool to ship features with, this is the control group, not the product.

prime-agent — the curator's take

The two abstractions are genuinely different: context as variables in a persistent REPL (not a transcript), and a harness that rewrites its own supplemental state with rollback — the closest thing to a shipping continual-learning agent. Daemon sessions and agent-to-agent messaging make it a long-horizon tool, not a chat CLI. When NOT: it's a full worldview — you adopt the RLM way or fight it; young codebase moving fast, and self-improvement means your harness drifts from everyone else's.