Foundations
Primitives
The raw materials every loop is assembled from — and what each one is for once the goal is autonomy instead of assistance.
The raw materials every loop is assembled from — and what each one is for once the goal is autonomy instead of assistance.
The hook
Ask ten engineers to "build a loop" and watch what they reach for. The weak ones start writing a scheduler from scratch. The strong ones open their agent's docs and discover that eight primitives — already built, already tested — snap together into the entire thing. Loop engineering is assembly, not invention. That reframe alone will save you weeks.
The eight primitives, seen through loop eyes (plain English)
The agentic coding primer met these as features you drive by hand. Here is the shift: each one maps cleanly onto a part of the loop's anatomy.
| Primitive | Hand-driven use | Loop use (the upgrade) |
|---|---|---|
| Plan mode | preview a risky change | rehearse a whole loop at L1 before granting L2 |
| Permissions | avoid annoying prompts | the L1→L3 ladder; the loop's hard boundary |
| Context | paste in the right files | keep each beat cheap; the spine is read first |
| Rules file | project conventions | the loop's constitution, loaded every beat |
| Skills | shortcuts for chores | the loop's trained moves, versioned in-repo |
| Hooks | format-on-save niceties | unbypassable guardrails between beats |
| Subagents | parallelize a big task | the maker/checker split that makes green mean done |
| MCP / connectors | query a database | the loop's hands: act on issues, PRs, messages |
Assembly example — the same tiny loop, twice
# Claude Code: a self-pacing report-only triage loop
# 1 rules file: CLAUDE.md says "report-only; one log line per run"
# 2 skill: .claude/skills/triage/SKILL.md holds the checklist
# 3 loop: /loop with no interval (self-paced), or Cron tools for schedules
/loop Run /triage. Append findings to state.md. Stop when the
inbox is empty, after 10 runs, or after 3 no-change runs.# OpenCode: same loop, cron-shaped
# 1 rules file: AGENTS.md carries the same two rules
# 2 permissions: opencode.json denies writes outside state.md
# 3 heartbeat: cron calls `opencode run "$(cat triage-prompt.txt)"`[!WARNING] Flags and file names drift weekly; the mapping (primitive → loop part) is the lasting layer. Cross-tool specifics live in the primitives matrix; commands live in each tool's docs.
[!NOTE] Going deeper: Part 3 (Day 2) gives worktrees, skills, connectors, and maker/checker a full lesson each. Sourced from Panaversity's Loop Engineering (S1) and Agentic Coding (S2), plus cobusgreyling/loop-engineering (S7).
Check yourself
Q: Your loop must (a) never touch secrets/ and (b) always summarize its diff in the PR
body. Which primitive carries each requirement — and why aren't they the same one?
Answer
(a) Permissions/hooks — it's a guarantee, so it has to be unbypassable machinery. (b) Rules file (or a skill) — it's a behavior, advisory by nature and perfectly fine that way. Put (a) in prose and it becomes persuadable; put (b) in a hook and you've built rigidity you'll come to regret. Guarantees in the harness, habits in the rules.
Try With AI
In your sandbox repo, ask your agent:
"Using only your built-in primitives — rules file, permissions, skills, hooks, subagents, MCP — sketch how you'd assemble a loop that keeps README.md's examples compiling. Name which primitive plays which part."
Grade its sketch against the table above: did it put the checker in a subagent? Did a guarantee quietly end up living in prose?
When it goes wrong
| Symptom | Cause | Fix |
|---|---|---|
| Loop reinvents scheduling/state in custom scripts | Assembly mistaken for invention | Use the tool's heartbeat + a plain state file first |
| Checker agrees with the maker suspiciously often | Same session graded its own work | Move the checker to a separate subagent (or separate loop) |
| Guardrail held for weeks, failed once at 3 a.m. | It was a rule, not a hook | Promote guarantees to permissions/hooks |
| Loop is powerful but terrifying | All eight primitives at L3 on day one | Climb the ladder: L1 → watch → L2 → watch → L3 |
Glossary terms used on this page: primitive, beat, spine, L1/L2/L3 — see glossary.md.
Sources: the primitives and their loop mapping come from Panaversity's Agentic Coding Crash Course (S2) and Loop Engineering: A Crash Course (S1), with assembly patterns from cobusgreyling/loop-engineering (S7). Full attribution: ../../resources/sources.md.