P—02 Case study · AI & ML
Conductor
Role
—
Status
active
Started
—
Last activity
Oct 2026
Stack
6 technologies
Conductor
Python · MCP · SSE
Overview
Conductor is a daemon that runs coding agents across 6 harnesses and 14 models. Every task gets a role, its own git worktree and an exclusive set of files it may write, so agents working in parallel never touch the same file, and a person approves every merge.
Prompts are split into a cached prefix that holds the role, the skill index and a ranked map of the repository, and a volatile tail that holds the task. That split kept prompt cache hits at 87-91% across live runs. The repository map is ranked by PageRank over the import graph.
Truncated output is detected from the structure of the agent's trajectory instead of by matching text, so a review that stops mid-sentence is never graded as complete. When a model refuses on quota, retries skip every model that shares that quota. More than 1000 tests cover its 13 modules.
Outcomes
Prompt cache hits
Tests
Models
Agent harnesses
Impact
- 01Runs coding agents across 6 harnesses and 14 models from one daemon, with a human approval gate before any merge
- 02Held prompt cache hits at 87-91% across live runs by splitting each prompt into a cached prefix and a volatile task tail
- 03Gives each task its own git worktree and an exclusive set of writable files, so parallel agents never collide
- 04Catches truncated agent output from the trajectory's structure, so a half-finished review is never graded as complete
- 05Covers 13 modules with more than 1000 tests
Features
01
One daemon, many agents
Runs coding agents across 6 harnesses and 14 models, each task assigned a role.
02
Isolated worktrees
Each task works in its own git worktree with an exclusive set of files it may write, so parallel agents never edit the same file.
03
Cached prompt prefix
Role, skill index and repository map sit in a cached prefix; only the task changes. Cache hits held at 87-91% across live runs.
04
Truncation detection
Cut-off output is caught from the structure of the trajectory, not by matching text, so an unfinished review never passes as complete.
05
Human merge gate
Nothing merges until a person approves it.
06
PageRank repository map
The repository map is ranked by PageRank over the import graph, so the files that matter most come first.
07
Quota-aware retries
When a model refuses on quota, retries escalate past every model that shares that quota.