A field guide

Loop Engineering

From prompting agents to designing the systems that prompt them.

Sentigent — the open-source loop that drives your coding agent to done. MIT · local-first

02

The 6 building blocks + memory

1

Automations

live
  • Scheduler that wakes on its own
  • Triages a backlog, re-fires the loop
  • Survives sessions
2

Worktrees

live
  • Isolated branches for parallel agents
  • Safe concurrent runs
  • operator/worktree.py
3

Skills

live
  • Reusable project knowledge
  • Steering file AGENTS.md · project-map.toml
  • closed-loop-delivery
4

Connectors / MCP

live
  • operator / loop / clone tools over MCP
  • Act inside real tools
  • gh review loop
5

Sub-agents

live
  • explorer / implementer / verifier
  • Maker–checker separation
  • Parallel fan-out (combined FAP)
6

Memory

live
  • External state on disk
  • ~/.sentigent/loops + local SQLite
  • episodes · rules · precedents · calibration
03

Context → Harness → Loop — the maturity model

Context → Harness → Loop
Advanced

Loop Engineering — design the system that prompts the agent

A loop carries the plan to done, then learns from outcomes. Reports one honest number, FAP (Faithful Autonomous Progress).

Plan→Act→Verify→Decide→Resume
Operating layer

Harness Engineering — the machinery the loop runs inside

The guardrails wall, the clone (push-vs-ask in your voice), MCP tools, worktrees for parallel runs, and the scheduler that re-fires it.

Foundation

Context Engineering — give it what you'd otherwise type by hand

The three layers — Profile · Plan · Policy — plus project-map and the auto-written steering file. The more context, the more it can safely do alone.

⚠Token costEvery lap spends tokens — tracked per run.
⚠Verification riskA silent pass lies — per-step verify + FAP.
⚠Comprehension debtCode you didn't write — project-map + diagrams.