Seven Commitments
Use this map to translate current loop-engineering vocabulary into seven design commitments for building safer, more useful AI work cycles.
Market vocabulary map
Current coverage and practitioner glossaries use terms such as agent loop, task loop, control loop, automations, worktrees, skills, connectors, sub-agents, cadence, kill switch, maker/checker, goal-condition primitive, and L1-L3 maturity tiers. The map connects them to the architectural commitments that make a loop safe and useful.
| Market term | How this site uses it | Book commitment / layer |
|---|---|---|
| Agent loop / task loop | A repeating AI work cycle that acts, observes, checks, and continues until a stop condition is met. | Scaffolding loop; full sensorimotor loop |
| Control loop | The older monitor-act-correct pattern that loop engineering adapts to AI agents, tool use, and review gates. | Full sensorimotor loop |
| Goal, check command, exit condition, max iterations | The operational shape of a well-formed loop: define done, run an objective check, stop on a named result, and cap runaway work. | Verifiable stop; materiality gating |
| Goal-condition primitive | An independent check that asks whether the loop's stated condition is actually true after an agent turn, rather than letting the acting model grade itself. | Verifiable stop; maker / checker |
| Automations / cadence | Scheduled or event-triggered loop runs, with frequency chosen by cost, risk, freshness, and response needs. | Materiality gating; visible verification |
| Worktrees | Isolated working directories or branches that let parallel agent work avoid file collisions and preserve rollback paths. | Motor + sensory; visible verification |
| Skills | Reusable capabilities or task instructions a loop can invoke instead of re-deriving project practice each run. | Long-term memory; debt accumulation |
| Connectors / plugins | Tool and API surfaces exposed to the agent through protocols such as MCP or platform-specific integrations. | Nervous system; motor + sensory |
| Sub-agents | Specialized secondary agents used for parallel work, review, synthesis, or independent verification. | Maker / checker; above-individual architecture |
| Memory / state | Durable files, trackers, or stores that let loop state survive context windows, crashes, and session boundaries. | Externalized memory; consolidation |
| Kill switch | A tested mechanism that halts an unattended or long-running loop before cost, permissions, or quality failure compounds. | Materiality gating; verifiable stop |
| L1 / L2 / L3 maturity tiers | Report-only, assisted patch-only, and unattended loop postures. The tier is an autonomy decision, not a quality score. | Materiality dial; human review |
Source anchors for this vocabulary include ADTmag's July 2026 coverage, the MCP.so loop-engineering guide, the public loop-engineering terminology repository, loopengineering.wiki, and current orchestration/governance tool comparisons.
Related engineering layers
| Term | Boundary |
|---|---|
| Prompt engineering | Optimizes a single instruction or interaction. |
| Context engineering | Chooses what files, memory, retrieval, and state the agent can see. |
| Harness engineering | Designs the tools, permissions, observability, and guardrails around the agent. |
| Loop engineering | Designs the repeated cycle: goal, state, tools, check, retry, escalation, and stop. |
The seven commitments
- Verifiable stop
- Maker / checker
- Externalized memory
- Materiality gating
- Frame before solve
- Visible verification
- Debt accumulation
Together, the commitments frame the work, preserve state, scale effort to stakes, separate making from checking, verify progress, stop deliberately, and record what the loop learns. Put them into practice with the prompts, worked examples, and living matrices.