
History of Self-Learning Harnesses and State of the Union in 2026
BlogSelf-learning harnessesHistory of Self-Learning Harnesses and State of the Union in 2026
以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。
BlogSelf-learning harnessesHistory of Self-Learning Harnesses and State of the Union in 2026
From the history of agent harnesses to ten prompts with Codex on cruise control.
A small Mac app that sits in the system tray and keeps prompting the Codex chats you choose. When Codex finishes, Cruise reads the conversation, figures out the next prompt, and sends it back to the same chat. You can also give Cruise guidance that evolves over time in Cruise.md.
We gave Cruise a revenue goal for a fictional launch site. The first four prompts left the goal unchanged. On prompt 5, Cruise saved the price and picked up a rule of its own.
November 7, 2026 · 10 Codex prompts · 5 edits to Cruise.md
An autopilot needs a way to decide what to do after each result. A feedback loop connects an attempt to an observation, then uses that observation to choose the next attempt. The useful signal might be a passing test, a better validation score, a reviewer’s correction, or evidence that a goal is still unfinished. The next prompt becomes useful when it responds to what happened.
In Andrej Karpathy’s autoresearch, an agent runs five-minute training experiments and keeps changes that improve a validation metric. Eigen Labs’ ECDSA.fail gives participants a public benchmark for comparing candidate quantum circuits. Anthropic’s automated alignment researcher uses Claude agents to propose experiments and iterate on whether a weaker model can supervise a stronger one. Each loop has a different way to judge the last attempt before deciding what to try next.
Two projects from 2025 show how far that feedback can reach. Google DeepMind’s AlphaEvolve generates code, tests it with automated evaluators, and uses the results to guide an evolutionary search for better algorithms. Sakana AI’s Darwin Gödel Machine modifies the coding agent itself, evaluates the resulting agents on programming tasks, and keeps an archive to explore further improvements. Both make the quality of the feedback central: another attempt becomes useful when its result can inform the next one.
In YC Paper Club’s Why the Harness Matters More Than the Model (7:13), François Chaubard traces how the machinery around a language model grew. His overview groups ideas by capability rather than strict chronology. These are some of its milestones, with dates checked against the original research.
His own autoresearcherUI brings several of these ideas together: experiments feed council reviews, lessons inform the idea queue, and a PI agent follows up on progress. The demo begins at 4:27.
We’re working on Cruise, which applies a narrower loop to existing Codex chats: read what happened, choose a useful next prompt, and retain guidance in Cruise.md.
Beyond the research milestones, the talk introduces tools that turn these ideas into everyday agent workflows.
RSI gives this history a larger ambition: improvements to a system help it make further improvements. In the strongest version, AI designs and develops its own more capable successors. That ambition has older roots. Jürgen Schmidhuber’s Gödel Machines (2003) described a theoretical system that rewrites its software after proving a change will improve its objective. The Darwin Gödel Machine later explored an empirical route, testing changes to an agent’s code. In 2026, Sakana AI’s MASS: Recursive Self-Improvement through Multi-Agent Self-Supervision studied a loop that improves team workflows, trains the shared model on selected executions, and uses the updated model in the next round.
It helps to name what is changing: the prompt, the memory, the harness around the model, or the model itself. Cruise’s automatic follow-ups and evolving Cruise.md operate at the prompt and memory level. That does not establish model-level RSI. In When AI builds itself, Anthropic distinguishes carrying out a human-supplied goal from choosing which goals are worth pursuing, and describes fully autonomous successor development as an open possibility. For every version of the loop, the practical question remains: what evidence shows that the change made the system better?
Earlier research explored how human guidance could support increasingly capable systems. In 2018, Iterated Amplification combined solutions to easier subproblems to build supervision for harder ones. In 2022, Anthropic’s Constitutional AI used written principles to guide self-critiques, revisions, and learning from AI feedback. They established useful ideas about oversight and durable guidance that help frame today’s agent loops.
The prompts themselves became something a model could improve. In 2022, Automatic Prompt Engineer (APE) generated candidate instructions and selected them using evaluated task performance. In 2023, Optimization by PROmpting (OPRO) put previous solutions and their scores into the next prompt, asking the model to propose better candidates. Applied to prompt optimization, this made the instruction part of the feedback loop too.
GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning (Agrawal et al., first submitted July 25, 2025) made this loop concrete for systems with one or more language-model prompts. GEPA stands for Genetic-Pareto. It reads execution traces, including reasoning, tool calls, and tool outputs, alongside evaluation feedback. A model reflects on what went wrong, proposes revised instructions, and tests those candidates. The search keeps prompts that work well on different examples, rather than repeatedly editing only one overall winner; it can also merge complementary improvements.
The underlying model weights stay fixed. What changes is the system’s written guidance, measured against a task evaluator. For this history, GEPA connects prompt optimization to the broader idea of a harness that learns from its runs. Its evaluation-driven search offers a useful comparison for evolving a saved prompt such as Cruise.md; editing that file alone does not establish the measured improvement that GEPA tests for.
In 2023, Voyager explored Minecraft by choosing new challenges, learning from execution feedback, and saving working code in a reusable skill library. That same year, Reflexion showed how an agent could write reflections on task feedback and carry those notes into later attempts. Together, they offer two useful ideas for an autopilot: choosing the next task and preserving something learned from the last one.
Truth Terminal grew out of Andy Ayrey’s 2024 Infinite Backrooms experiment, where language models had open-ended conversations with one another. Ayrey trained Truth Terminal on related conversations and gave it a Twitter account. Its posts drew a following and inspired meme coins created by other people. For me, its legacy is that it made agent feedback visible in public. People could react to what the agent posted, and those reactions raised questions about who stewards an agent and the resources that accumulate around it. Truth Collective’s account
OpenClaw started as Peter Steinberger’s WhatsApp relay and went through the names Clawd and Moltbot before becoming a personal agent platform that runs on your machine. It keeps durable decisions in MEMORY.md and working notes in dated Markdown files. Nous Research’s Hermes Agent also keeps memory across sessions and can turn a workflow it figured out into a reusable skill. Those files give the next session something concrete to build on. OpenClaw’s origin · OpenClaw’s memory · Hermes memory and skills
In Karpathy’s LLM Wiki, an agent keeps the source material and maintains a separate set of linked Markdown pages. New evidence can change an existing page or surface a contradiction. On Hacker News, the maker of WUPHF described a related approach that uses Markdown and Git for a shared wiki. Agents draft notes and reviewers decide what belongs in the wiki, with Git recording the edits. LLM Wiki · WUPHF on Hacker News
Google’s Elephant-Goldfish Model uses a long-running session, the elephant, to draft a specification. Fresh sessions, the goldfish, review the draft and its references without seeing the earlier chat. They can spot assumptions that made sense in the conversation but never made it onto the page. Elephant-Goldfish Model
Memory also needs a way to outlast the current context window. MemGPT explored that in 2023 by letting an agent move information between memory tiers, drawing on ideas from operating systems. The work continued through the Letta agent framework. For Cruise, the relevant lesson is continuity: a later prompt should be able to use the goal and useful lessons saved in Cruise.md.
These ideas now appear in everyday products. Cursor’s Background Agents (2025) brought asynchronous coding to remote environments; its Automations (2026) launch cloud agents on schedules or events. Grok Bot’s skills and routines (2026) turn a successful workflow into reusable instructions, then run it on a schedule or event. Codex Goals (2026) keeps an objective active across turns until it reaches a supported stopping condition. Claude Code’s /goal (2026) checks a completion condition after each turn and continues when it is unmet. Cruise works at the Mac menu bar, following selected Codex chats and carrying their guidance in Cruise.md.
Another example is Vorflux, from Rippling cofounder and former CTO Prasanna Sankar. Its engineering harness breaks work into tasks and repeats a build, judge, and fix loop. On November 5, 2026, Sankar announced Vorflux 2.0 as Slack coworkers that own a goal and work around the clock. That announcement points toward agents carrying an ongoing responsibility, with the product’s public documentation providing more detail on its engineering and review loops.
YC’s Summer and Fall 2026 directories show more companies building the machinery around an agent. Jcode’s Auto-poke checks unfinished work and prompts the agent to continue. Its self-development mode can edit, build, test, and reload its own implementation. Hoplite coordinates cloud coding agents and brings them back to CI failures and review comments. Acyclic Labs splits work into sandboxed agent branches and combines their results. These are close relatives of the orchestration offered by Conductor and Orchestrator: keeping work moving across agents, tasks, and checks.
A related group concentrates on what the next run can learn. Glen retrieves prior sessions and distills shared skills. Grep AI describes a self-improving harness that turns repeated reasoning steps into workflows and code. Papaya proposes fixes from agent traces and monitors the results. For research itself, Autoresearch, Inc. runs evaluated improvement campaigns, while rekursiv.ai aims to use AI scientists’ discoveries to accelerate their next experiments. The shared opportunity is to make experience useful after the current task ends.
Cruise applies that feedback pattern to an existing Codex chat. The completed reply helps Cruise choose the next prompt, and a saved goal in Cruise.md gives those prompts a direction. Submitting a prompt sends a specific task to Codex, like “Add a sponsorship inquiry form.” Saving a goal gives Cruise guidance it can reuse across prompts, like “Explore revenue while keeping ordinary submissions free.” The checkbox controls which chats get automatic follow-ups.
PromptWork +questionsReview + guidanceNextpromptEvery reply informs the next prompt.
The idea came from a recurring bottleneck: models were getting remarkably good, and running several agents in parallel was becoming practical. A frontier lab subscription with enough capacity, or a workspace for orchestrating agents such as Conductor, could keep more work moving at once. Human attention was often the thing holding it up.
Each reply still needed review, another prompt, an approval, an answer, or a nudge back toward the intended direction. Those small interventions add up. More of that attention belongs on direction and judgment, with less time spent typing the next useful follow-up.
Tha
正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。