Skip to content
/
Phase B Batch 5

Batch 05 Notes — Major Coding Agents (Terminal/IDE Harnesses)

Batch 05 Notes — Major Coding Agents (Terminal/IDE Harnesses)

Agent: Phase B batch agent #05 of 33
Theme: Major coding agents (terminal/IDE harnesses)
Frameworks analyzed: aider, cline, opencode, goose, continue, tabby, gpt-engineer, openhands, swe-agent, sweep
Date completed: 2026-05-26

Roster Table

Framework Stars License CLI Binary Local UI Orchestration Multi-Model Tier
opencode 165,457 MIT opencode desktop-app task-decomposition-tree yes A
openhands 74,904 NOASSERTION openhands web-dashboard sequential no A
gpt-engineer 55,211 MIT gpte none sequential no C (archived)
cline 62,338 Apache-2.0 cline vscode-extension parallel-fan-out no A
goose 45,882 Apache-2.0 goose desktop-app hierarchical no A
aider 45,340 Apache-2.0 aider terminal-tui sequential yes A
tabby 33,542 NOASSERTION tabby web-dashboard none yes A
continue 33,391 Apache-2.0 cn vscode-extension sequential no A
swe-agent 19,314 MIT sweagent none sequential no A
sweep 7,712 NOASSERTION sweep none (JetBrains plugin separately) sequential no A (dormant)

Total stars: ~537,000 across 10 frameworks

Intra-Batch Patterns

  1. Every framework ships a CLI binary — all 10 have a named binary (aider, cline, opencode, goose, cn, tabby, gpte, openhands, sweagent, sweep). This contrasts with seed frameworks (BMAD, spec-kit) which are methodology docs without executables.

  2. Isolation bifurcation: 3 frameworks run in Docker containers (Tabby, OpenHands, OpenCode via Docker support), 7 run in-place. Docker users score better on SWE-bench because they can safely run test suites and undo file changes. In-place runners avoid Docker dependency but risk polluting the host environment.

  3. The REPL-vs-pipeline split: SWE-agent forces one command per turn (REPL loop); all other frameworks allow multi-step internal reasoning before producing output. SWE-agent's REPL constraint is a research design choice for trajectory analysis; it would be inconvenient in production.

  4. Multi-model routing is rare but significant: Only aider (Architect/Editor) and opencode (per-agent routing) implement true per-role model assignment. The majority (8 of 10) use a single model throughout.

  5. Pivot stories are common: 2 of 10 frameworks show clear product pivots — gpt-engineer (archived, spawned Lovable.dev), sweep (GitHub bot → JetBrains plugin). The pattern: GitHub bot → IDE plugin as the commercial destination.

Most Interesting Find

Sweep's sweep.yaml blocks its own prompt file: The Sweep repository's own sweep.yaml includes blocked_dirs: ["sweepai/core/prompts.py"], meaning Sweep (the GitHub bot) is configured to refuse to modify its own core system prompts when processing issues against the Sweep repo. This is a self-referential meta-constraint that demonstrates the sweep.yaml behavioral rule injection taken to its logical extreme. No other framework in this batch has a comparable self-referential constraint in its own config.

Tier C Items

  • gpt-engineer (slug: gpt-engineer): Archived repository (last push May 2025, ARCHIVED status). Sufficient material for full 11-file analysis — the preprompts/ directory contains well-documented prompt philosophy. The gpte binary still installs but the project is officially discontinued. Marked maintainer_status: archived in METRICS.yaml. Documented as the direct precursor to Lovable.dev (web app generation) and spiritual ancestor of all single-shot code generation frameworks.

Status Notes

  • sweep: Marked maintainer_status: dormant (last commit 2025-09-18, no active development on the open-source repo). The JetBrains plugin at plugins.jetbrains.com/plugin/26275-sweep-ai is active but is a separate product not in this repo. All 11 files cover the open-source GitHub bot/CLI codebase.
  • tabby: Contributor count capped at 100 (GitHub API max_per_page); actual count may be higher.
  • opencode: Repo moved from sst/opencode to anomalyco/opencode during analysis. All API calls resolved via the original URL (GitHub redirect) and canonical reference included in files.
  • openhands: agenthub/ directory returned 404 (structure changed in main branch). Analysis used .openhands/microagents/ and skills/ directories which contain the equivalent content.

Cross-References Discovered

  • aider SEARCH/REPLACE format is referenced or adopted by: opencode (same format), SWE-agent (edit_anthropic tool bundle), cline (partial)
  • SWE-bench benchmark is the common evaluation target for: openhands (77.6%), SWE-agent (original benchmark developers), sweep (unmeasured, predates SWE-bench)
  • sweep.yaml rules pattern is the ancestor of: .clinerules/ (Cline), .cursorrules (Cursor), CLAUDE.md (Claude Code) — all implement the same "project rules as behavioral constraints" paradigm
  • gpt-engineer's preprompts/philosophy (coding principles as system prompt constraints) directly anticipates OpenHands' microagent PERSONA pattern and Sweep's aspirational quality claims
  • Docker two-container architecture (OpenHands: runtime + sandbox) is a more sophisticated version of the container isolation that tabby uses for server deployment
← Batch 4
Batch 6 →