Evidence-linked trend hypothesis
Goal-scoped loops replace manual continuation
Long-running agent work is moving from repeated continue prompts toward explicit goals, stopping conditions, schedules, task state, and evaluator-driven retries.
Current thesis
An autonomous loop becomes operationally useful when completion is an external contract rather than a model's subjective decision to stop. Confidence: high. Supported by 140 normalized raw signals. This New Runtime record is an evidence-linked retrieval unit. Use its canonical page, machine-readable representations, dates, scope, and public source URLs to verify the claim before reusing it.
Source ledger
Publishable sources attached to this record.
| # | Source | Role | Public status |
|---|---|---|---|
| 1 | code.claude.comsource | primary receipt | source_urls |
| 2 | github.comrepo | supporting receipt | source_urls |
| 3 | cursor.comsource | supporting receipt | source_urls |
| 4 | github.blogrepo | supporting receipt | source_urls |
| 5 | websitemain.claude.comsource | supporting receipt | source_urls |
Showing 5 of 7; the complete set is exposed in the JSON route.
Current state, as of Oct 13, 2026
Where this stands now
strengthening The pattern is still holding: goal-scoped loops use evidence outside the acting model to decide whether work is done. A 2026-10-05 update describes Gauntlet Loop measuring repeated improvement against a concrete quality reference.
Goal-scoped loops remain useful when an external condition defines completion, rather than the model deciding it is finished. The Gauntlet Loop is a concrete example: repeated improvement is measured against a quality reference; other recent signals describe persistent parallel work and deliverable validation.
The evidence strengthens the pattern but does not show that every loop needs the same structure. Loops suit repeated work with one stopping rule; graphs add explicit branches, shared state, parallel work, approvals, and recovery when exceptions matter. Builders should define the target and evaluator, and add graph structure when task routes require it.
What changed since the original text
- A Gauntlet Loop update describes agent improvement measured against a concrete quality reference rather than the agent's own sense of completion. 12
- September updates cite persistent parallel work and documented stopping points, state-bearing sessions, and deliverable validation as further examples of external completion contracts. 3
- An archive item distinguishes simple repeated loops from graphs for branching, parallel work, approval gates, shared state, and recovery; it frames graphs as an extension, not a replacement. 4
Open questions
- How often do deployed loops use independent evaluators rather than model-generated judgments?
- When do added graph controls improve reliability enough to justify their complexity?
Watch next: Watch for deployment evidence showing whether external evaluators and explicit recovery paths improve outcomes on long-running agent tasks.
- Weekly update: Gauntlet Loop uses an external quality reference for agent improvement update, Oct 5, 2026
- The Gauntlet Loop: The Prompting Method Behind Claude of Duty archive
- Weekly update: New tools and docs show loops anchored to external contracts become operationally useful update, Sep 21, 2026
- Agent Loops Become Graphs When Exceptions Matter archive
Rewritten automatically by the New Runtime knowledge desk from this record's weekly updates, releases and archive. The original text below is kept as written (May 14, 2026); where the two disagree, this section is newer.
Weekly knowledge desk
Weekly updates
3 reviewed updates to this pattern, newest first. Every update is checked against its evidence before it changes the record.
Habr author describes AIWorks using bounded, verifiable agent loops
The author describes a local supervisor for Claude Code where unattended work depends on a bounded loop with verifiable outputs, rather than simply asking the model to continue. That is a concrete example of the record’s goal-scoped completion contract, with a human-controlled merge boundary.
Gauntlet Loop uses an external quality reference for agent improvement
The Gauntlet Loop description says repeated agent improvement is measured against a concrete quality reference rather than the model's own decision that it is finished. That is a specific instance of the record's goal-scoped completion contract.
- The Gauntlet Loop: The Prompting Method Behind Claude of Duty signal, Oct 10, 2026
New tools and docs show loops anchored to external contracts become operationally useful
Anthropic’s Claude Projects redesign (e133, e75) lets coordinators delegate work to parallel threads that persist beyond a single conversation, and OpenAI’s documentation (x1, x2) explicitly defines loop continuation around external stopping points, state‑bearing sessions, and deliverable validation. These independent signals demonstrate that autonomous loops are being built with goal‑scoped completion contracts rather than relying on a model’s internal sense of finish, reinforcing the record’s claim.
- Projects redesigned: from folder to conversation signal, Sep 23, 2026
- Claude rolls out Projects in Claude Code with parallel background threads signal, Sep 24, 2026
- Running agents | OpenAI API external, Sep 28, 2026
- Run long horizon tasks with Codex | OpenAI Developers external, Sep 28, 2026
What replaces keep going
Repeated continuation leaves the stopping rule inside the same model that is doing the work. Goal-scoped systems move it into observable state: an acceptance test, score, deadline, retry budget, queue state, or approval. The agent can then continue, recover, or stop for a reason the operator can inspect.
Supporting evidence
- Claude Code goals bind work to an explicit outcome.
- Maintainer scripts package recurring repository work as stateful loops.
- Cursor and GitHub turn successful interactive workflows into automations.
- Continuous evaluation gives multi-agent systems an external quality signal.
Operational consequence
Every loop should declare its target, evaluator, budget, permissions, retry policy, checkpoint, and escalation path. Tasks without a credible evaluator should remain interactive rather than receiving artificial autonomy.
September 2026 update
What changed
Conversation-shaped projects, inspectable long-running trajectories, and a managed deep-agent terminal show stateful work being operated as a task with a boundary.
Builder rule
Specify acceptance evidence, budget, allowed tools, checkpoint cadence, escalation owner, and stop/rollback condition before the loop begins.