Manual continue prompting → goal-scoped agent loops
Long-running agent work is moving from manual continue prompts toward explicit goals, limits, subagents, schedules, and verifiable completion.
Tectonic shift
Before: Human repeatedly nudges one long chat forward After: Agent loop runs against an explicit goal, budget, checks, and stop condition Current stage: emerging 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 | code.claude.comsource | supporting receipt | source_urls |
| 3 | code.claude.comsource | supporting receipt | source_urls |
| 4 | langchain.comsource | supporting receipt | source_urls |
| 5 | websitemain.claude.comsource | supporting receipt | source_urls |
| 6 | github.comrepo | supporting receipt | source_urls |
Current state, as of Oct 9, 2026
Where this stands now
strengthening Major AI platforms now embed explicit goals, budgets, verification steps, and stop conditions into agent loops, moving work from manual continuation prompts to autonomous, goal‑scoped execution.
As of October 2026, LangSmith, Warp, and Anthropic have released production‑grade tools that require agents to declare an objective, a budget, acceptance checks, and a termination condition before execution. These platforms also provide managed deep‑agent runtimes that persist state across sessions and enforce deterministic promotion of successful runs.
The shift is reflected in deployments such as Google Managed Agents and OpenAI Workspace Agents, where agents operate under schedules, permissions, and resumable identities. The tooling now automates the outer loop—changing prompts, tools, and routing—while retaining deterministic evaluation, reducing the need for manual nudging.
For builders, this means designing agents as components of a controlled loop rather than as stand‑alone chat continuations. Workflows must specify goals, resource caps, and verification rubrics up front, and integrate with the platform’s evaluation APIs to achieve reliable, long‑running automation.
What changed since the original text
- New tools and workflows embed explicit goals, budgets, and self‑evaluation into AI agent loops (LangSmith, Warp, Anthropic). 1
Open questions
- How will standardized evaluation rubrics evolve to handle domain‑specific success criteria?
- What mechanisms will ensure robust recovery when an agent loop fails mid‑execution?
Watch next: A release that adds automated rollback and state‑checkpointing to managed agent runtimes would further solidify the goal‑scoped loop model.
- Weekly update: New tools and workflows embed explicit goals, budgets, and self‑evaluation into AI agent loops update, Sep 21, 2026
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 (Jul 21, 2026); where the two disagree, this section is newer.
Weekly knowledge desk
Weekly updates
3 reviewed updates to this shift, newest first. Every update is checked against its evidence before it changes the record.
Durable agent harnesses and bounded loops surface in last week’s coverage
Several items describe work organized around persistent or bounded agent processes rather than repeated continuation prompts. Claude Code and Vercel are described as supporting durable or resumable jobs, while an AIWorks account emphasizes explicit loops, verifiable outputs, and a human-controlled merge boundary; OpenAI’s account of Third Moment Research likewise places long-running work in the harness. These are concrete, independent examples aligned with the record’s shift toward goal-scoped agent loops.
Record state: confidence high.
- Long-running agent work is becoming durable infrastructure, not a chat session signal, Oct 24, 2026
- An author on Habr describes building AIWorks, a local supervisor for Claude Code post, Oct 21, 2026
- How Third Moment Research runs 10,000 agents for 36+ hours post, Oct 19, 2026
Gauntlet Loop uses goals and critic agents for repeated improvement
Matt Shumer’s Gauntlet Loop describes giving a lead agent a goal and a concrete quality reference, splitting the work, and having critics assess results against that reference. This is a concrete example of agent work organized around a goal and verifiable checks rather than repeated nudges in one long chat.
- The Gauntlet Loop: The Prompting Method Behind Claude of Duty signal, Oct 10, 2026
New tools and workflows embed explicit goals, budgets, and self‑evaluation into AI agent loops
Recent announcements from LangSmith, Warp, and Anthropic describe platforms that turn agent operations into closed improvement loops with defined objectives, budget limits, and verification steps. LangSmith’s managed deep agents and trajectories aim to automate goal‑scoped execution, while Warp Factories adds rubric‑based scoring and self‑improvement agents that adjust setups based on performance. Anthropic’s Claude Code Projects redesign introduces goal‑driven coordination spaces that delegate work to parallel sessions with persistent context. Together these independent developments provide concrete instances of the shift from manual prompting to structured, goal‑oriented agent loops.
Record state: stage accelerating, confidence high.
- 1️⃣ LangSmith Engine v2 2️⃣ Managed Deep Agents v0.8 3️⃣ LangSmith Trajectories 4️⃣ LangSmith Fine-Tuning 5️⃣ Custom Apps Everything we announced at Interrupt NYC ⤵️ https://t.co/8cnORN5IbT post, Sep 26, 2026
- Warp adds rubric-based scoring and self-improvement agents to Warp Factories post, Sep 24, 2026
- Projects redesigned: from folder to conversation signal, Sep 23, 2026
A long-running agent does not become more reliable when the user says “keep going”. It needs a goal, budget, acceptance checks, resumable state, and a visible stop condition.
Evidence
The related pieces on Claude Code loops, cost controls, and AI-native team stages show the same pattern: agent autonomy is moving from the prompt into the operating model.
Bottleneck
The main bottleneck is not the model, but controllability: which actions are allowed, where the human confirms the result, how the loop knows that work is finished, and how the system recovers after failure.
September 2026 update
What changed
Trajectory review, project-based parallel work, and a managed deep-agent terminal give long-running work a visible lifecycle.
Current boundary
Define objective, budget, artifact, check, owner, and stop condition. A loop without these remains a continuation prompt with more compute.