aisettings
Log InAdd your setup

Mitchell Hashimoto

Last updated: Sep 14, 2026
RoleDeveloper; co-founder of Superlogical
Sources16 cited fields · high confidence
@mitchellhmitchellh.com

Uses Amp as his main agentic coding tool, plans with its oracle, and keeps manual review and harness engineering in the loop.

“Final manual review is super super super important.” mitchellh.com

Share cardShare your setupCompare

Loadout

aisettingsShare on X

Mitchell Hashimoto’s loadout

Harness
Amp
His October 2025 Ghostty feature write-up calls Amp the agentic coding tool he uses most.
Model
GPT-5.5
xhigh
Also uses
Claude Fable
Runs at xhigh as planner/architect and judge.
TerminalSuperlogical multiplexer (Rex)
ToolsSuperlogical CLI+2 more

3 views · 0 equipped

Models

GPT-5.5xhigh
Uses GPT-5.5 at xhigh as the coder in a Fable-planner/Fable-judge pipeline.
x.com, 2026-07 ↗
Claude Fable
Runs at xhigh as planner/architect and judge.
x.com, 2026-07 ↗
GLM-5.1
Head-to-head comparison model for iterative feature/change work.
x.com, 2026-06 ↗
Gemini
Used to reproduce a Zed command palette in SwiftUI; the result lightly influenced Ghostty's shipped macOS palette.
mitchellh.com, 2026-02 ↗

Editor & terminal

Superlogical multiplexer (Rex)
His current terminal-multiplexer work includes a CLI for sessions, commands, splits, input simulation and agent-tool hooks.
x.com, 2026-09 ↗

Tools

Superlogical CLI
Controls the multiplexer and exposes streaming events plus terminal process information for automation.
x.com, 2026-09 ↗
Amp oracle
Read-only, slower, higher-cost subagent used for planning.
mitchellh.com, 2025-10 ↗
AGENTS.md and programmed verification tools
Encodes fixes for recurring agent mistakes and adds scripts such as screenshots and filtered tests.
mitchellh.com, 2026-02 ↗

Workflow

  1. 1Plan with Fable xhigh, code with GPT-5.5 xhigh, then have Fable xhigh judge the result.x.com, 2026-07
  2. 2Share plans by copy and paste and personally read the code to check the result.x.com, 2026-07
  3. 3Plan manually before invoking the agent, then ask for a plan-only session on vague or non-trivial work.mitchellh.com, 2025-10
  4. 4Save comprehensive plans as spec.md and ask later sessions to consult the spec for focused tasks.mitchellh.com, 2025-10
  5. 5Use end-of-day agents for research, parallel idea exploration, and issue or PR triage.mitchellh.com, 2026-02
  6. 6Clean up and manually review generated code; he calls this an anti-slop session.mitchellh.com, 2025-10

Timeline

  1. 2026-09-14TerminalSuperlogical multiplexer (Rex)x.com, 2026-09
  2. 2026-09-14ToolsSuperlogical CLIx.com, 2026-09
  3. 2026-07-02ModelGPT-5.5x.com, 2026-07
  4. 2026-07-02Effortxhighx.com, 2026-07
  5. 2026-07-02Also usesClaude Fablex.com, 2026-07
  6. 2026-06-10Also usesGLM-5.1x.com, 2026-06
  7. 2026-02-05Also usesGeminimitchellh.com, 2026-02
  8. 2026-02-05ToolsAGENTS.md and programmed verification toolsmitchellh.com, 2026-02
  9. 2025-10-11HarnessAmpmitchellh.com, 2025-10
  10. 2025-10-11ToolsAmp oraclemitchellh.com, 2025-10

Comments

Log in to comment