新智元报道
单Agent最好的一次，找出27个Bug。
换成Claude自己组队，连跑三次，每次都查出66个。
这是Anthropic刚公布的一组测试结果：在11.6万行代码中预先埋入70个Bug，让单Agent和动态工作流各跑三次。
结果，单打独斗，最好也没查出四成；而组队协作，三次都查出了九成以上。
这组结果背后，是Claude开始自己安排分工、组织多个Agent协作。
10月10日，Claude Managed Agents动态工作流开启公测。
有了这项功能，主Agent接到任务后，可以自己编写分工程序，安排其他Agent分阶段完成任务，最后汇总结果。
同一天，Claude Code Projects也扩大了公测范围：所有此前加入等待名单的Pro和Max用户，现在都已获得使用资格。
在Projects里，你可以围绕一个项目持续提出需求，Claude负责拆任务、协调多条线程并行推进。
一个让开发者把多Agent协作接进自己的应用，一个让用户在项目对话里直接交代工作、跟进进展。
两项更新指向同一个变化：除了执行任务，分活、盯进度、交接结果，这些原本需要人来操心的事，Claude也开始接手了。
同时开几个AI窗口不难。
麻烦的是，给每个窗口重讲一遍背景，把一边的发现送到另一边，再挨个追问：做到哪了，卡在哪里，还缺什么？
窗口越来越多，人反而忙成了传话员。
Projects想接下这部分工作。
你在同一个项目对话里持续提出需求，Claude判断该新开一条任务线程，还是交给已经在处理相关工作的线程。
每条线程可以理解为一段独立推进任务的工作对话。
Projects协调对话与任务总览
Anthropic举过一个例子：让API、网页端和移动端一起停用旧接口。
这件事涉及多个代码仓库，各处修改还得互相配合。
Claude可以按代码仓库拆出任务，分别修改、运行测试、提交代码合并请求，再告诉你哪些改动需要先合并。
每条云端线程有自己的上下文和代码副本，在自己的分支上工作，再把结果报回项目对话。
你仍能进入某条线程看细节、纠正方向。项目总览则把正在工作、等待你回答、可以审查的任务分开列出。
如果离开一会儿再回来，你也不必挨个窗口问进度，可以直接从需要你处理的地方接上。
要省下反复沟通的时间，之前交代过的事就必须记得住。
Projects会积累项目记忆。
需求改了什么、做过什么决定、哪些地方踩过坑，都可以记录下来，供后续云端线程读取。
官方举了一个日常工作场景：发布日期改到了周五，某项功能为什么被砍掉，修改某个服务前要先找谁确认，这些信息都可以留在项目记忆里。
资料库也会保存上传的资料和Claude生成的文件，让后续任务接着已有成果往下做。
需求分配到右侧线程，相关决定和工作结果逐步积累，供后续任务使用。
新版Projects在9月17日已开启分批公测。
这次扩大的是准入范围，功能本身仍处于公测阶段，尚未报名的用户仍可加入等待名单。
云端线程可以在合上电脑后继续工作。需要本机工具或本地数据库的任务，也能通过Remote Control在电脑上运行，但电脑必须保持唤醒。
项目有人协调之后，一项复杂任务内部又该怎样分工？
Managed Agents动态工作流，处理的就是这类问题。
官方给出了一个具体任务示例：
检查300份合同，找出哪些包含控制权变更条款，也就是公司控制权发生变化时，合同该如何处理的约定。
逐份读懂已经很费工夫，还得确保每份都查过、结论有据可查，漏掉的也能及时补查。
Claude会为任务写出一段工作流程序，安排哪些Agent读材料、哪些步骤处理结果，以及后续怎样继续。
合同阅读任务并行展开，再进入核对与汇总
分工和交接都写进了程序，可以直接运行。
一个Agent读完材料，结果交给程序。程序再把结果传给后续Agent，或者据此决定下一步走哪个分支。
需要反复修改的工作，也能把循环安排进去。
官方文档举的例子是稿件审校：持续修改，直到审核通过，或者达到预设轮数。
普通子Agent委派中，主Agent要读汇报、接着决定下一步。动态工作流则把大量中间交接写进程序，在后台推进。
等待期间，主Agent仍可以和用户交流、查看进度。运行结束后，再读取结果，答复用户或发起下一轮工作。
不过，工作流中的线程交回结果后，主Agent不能像调用普通子Agent那样，继续向同一线程追问。因此，核对和返工最好提前安排进流程。
官方的合同审查指令样例就规定：
第一轮，每份合同交给一个Agent；第二轮，由另一个Agent重新检查全部合同，连首轮未发现目标条款的也不能跳过；复核未通过的，返工后再次复核。
这给首轮漏掉的条款，多留了一次被发现的机会。
当然，合同案例展示的是工作方式。
Anthropic并未在那组Bug测试帖子中披露具体分工，不能据此认定，这66个Bug也是通过同样的流程找出来的。
单次最多1000个Agent，是这次公测最抢眼的数字之一。
它指的是一次工作流在整个运行期间，累计最多启动1000个Agent。
当前文档列出的同时工作线程上限为64，官方也说明，这个并发数可能调整。
工作流可以一批接一批地执行，也可以等前一阶段交回结果，再安排后一阶段。
每个Agent有独立的对话历史，同时共享会话中的文件和沙箱，也就是运行代码、处理材料的工作环境。
开发者可以提前配置专门的Agent，也可以让工作流根据任务需要临时定义。
接入时，在multiagent配置中将type设为multiagent_20261001，再配置模型、工具和任务要求。
启用动态工作流后，向Claude交代任务，便可让它编写分工计划。图中示例为筛查300份合同中的控制权变更条款。
不过，会分工还不够，还得把没做完的地方交代清楚。
官方的一条指令样例要求：某个Agent读不了合同，就把这份合同列为「未覆盖」，其余任务继续。
「没查出问题」和「根本没查到」必须分开。否则，一份整齐的汇总，很容易把任务缺口藏起来。
开发者可以查看运行阶段和各条线程的记录，沿着记录定位问题。
以合同审查为例，开发者可以跟踪阅读、核对等阶段，并查看每条任务线程的执行记录。
Agent多了，既要安排好执行顺序，也要看清哪些任务完成了、哪些仍有缺口。
需要区分的是：最多启动1000个Agent、三次均查出66个Bug，说的都是Managed Agents动态工作流，并非Projects。
回到开头的测试：单Agent三次找出14、15、27个Bug，动态工作流三次都是66个。
查出的Bug更多了，但为此多花了多少时间和费用，还不清楚。
官方没有披露所用模型、实际Agent数量、耗时、Token用量和误报情况，暂时还无法判断这套工作流是否更快、更划算。
实际使用时，两项功能的费用也要分开看。
Projects使用现有套餐额度，工作线程和协调对话都会消耗用量。并行任务越多，额度也会用得越快。
用户可以调整模型和推理强度，或让Claude少开一些线程。通常，工作线程达到套餐用量上限后会暂停，等额度重置再继续。
Managed Agents则按模型Token用量计费。此外，每个会话每小时另收0.08美元运行费，只计算会话处于运行状态的时间。
这0.08美元只是运行费，多Agent阅读材料、生成答案消耗的Token还要另算。
开发者可以设置会话预算，工作流的模型消耗也计入其中。达到预算后运行暂停，但已发出的模型请求仍会完成，最终费用可能超出预算。
会话预算覆盖动态工作流中的模型消耗，方便控制多Agent协作的支出。
因此，官方建议先从范围明确的任务开始，摸清消耗，再逐步增加复杂度。
Agent多了，并不自动等于更划算。
多找出多少真实问题，增加了多少模型调用费用，又需要人花多少时间复核、修错，都得算进同一笔账。
少花时间分活，如果换来更多收尾工作，项目未必更快结束。
这两项更新，让Claude开始承担更多组织工作。用户交代目标、预算和验收标准，Claude负责分工、推进和交接。
接下来的考验是：把活分出去之后，Claude能不能少让人操心，把经得起检查的结果交回来。
参考资料：
https://x.com/ClaudeDevs/status/2108591328732856655
https://x.com/ClaudeDevs/status/2108591330129523146
https://x.com/ClaudeDevs/status/2108591331660468538
编辑：元宇 摩西
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
One of Agent's best. Find 27 Bugs.
Claude's own team ran three times, 66 each time.
This was a set of tests just released by Anthropic: 70 Bugs were pre-positioned in 116,000 line codes, allowing single Agent and dynamic workflow to run three times.
As a result, 40 per cent had not been identified either, and teams had worked together to identify more than 90 per cent on three occasions.
Behind this set of results is Claude beginning to organize his own division of labour and organize multiple Agent collaborations.
On 10 October, Claude Managed Agents Dynamic workflow opened public surveys.
With this function, Major Agent, upon receiving a mandate, can prepare a division of labour process for himself, arrange for other Agent to complete the task in stages and eventually consolidate the results.
On the same day, Claude Code Projects also expanded the scope of the survey: all Pro and Max users who had previously joined the waiting list were now eligible to use it.
In Projects, you can continue to demand around a project where Claude is responsible for dismantling tasks and coordinating multiple lines in parallel.
One allows developers to connect multiple Agent to their own applications, and one allows users to speak directly to and follow up on progress in project dialogues.
Two updates point to the same change: in addition to the task, division, tracking of progress, handover of results, which would have required attention, Claude began to take over.
It's easy to open several AI windows at the same time.
The trouble is to rephrase the background to each window, send the discovery of one side to the other, and then ask the question: Where are we, where are the cards, what are we missing?
More and more windows, people become messengers.
Projects want to take over this part of the job.
You continued to demand in the same project dialogue, and Claude judged whether the opening of the new task line was to be given to work already done.
Each thread can be understood as a working dialogue to advance the mandate independently.
Projects coordinate dialogue and task overview
Anthropic has given an example: to disable the old interface with API, the web page and the mobile side.
The matter involved multiple code warehouses, and changes had to be made to one another.
Claude can remove tasks by code repository, modify, run tests, submit code consolidation requests separately, and then tell you which changes need to be consolidated first.
Each cloud-end thread has its own context and code copy, working on its own branch and reporting back to the project dialogue.
You can still go into some line to see the details and correct the direction. The project overview lists separately the tasks that are in work, are waiting for your answer and can be reviewed.
If you leave later, you don't have to go through a window to ask for progress, you can go straight to where you need to handle it.
In order to save time for repeated communication, previous accounts must be kept in mind.
Projects accumulate project memory.
What needs to be changed, what decisions have been made and where the pits have been stepped can be recorded for subsequent cloud line reading.
Officially, there is a day-to-day working scene: the date of publish has been changed to Friday, why a function has been cut off, and who needs to be identified before modifying a service, and this information can be left in the project memory.
The database also saves uploaded information and documentation generated by Claude, allowing follow-up missions to proceed with results already achieved.
Demand is assigned to the right-hand line and decisions and results are cumulative for subsequent missions.
The new version of Projects has already started a series of public surveys on 17 September.
This expanded access, the functionality itself is still in the public domain and users who have not yet registered can still join the waiting list.
Cloud-end threads can continue to work once the computer is installed. Tasks that require an in-house tool or a local database can also be operated on the computer through Remote Control, provided the computer is awakened.
What would be the division of labour within a complex mission once the project was coordinated?
These are the issues addressed by Managed Agents Dynamic workflow.
An example of a specific mission is given:
Check 300 contracts to find out which agreements contain a change of control clause, i.e., how the contract should be treated in the event of a change in the company's control.
Read-by-reading is already very difficult, and it is also necessary to ensure that each of them is checked, that the conclusions are documented and that missing items can be followed up in a timely manner.
Claude will write a workflow program for the mission, which Agent reading materials will be arranged, which steps will be taken to process the results, and how to proceed.
The contract reading missions were carried out in parallel, and then entered into reconciliation and aggregation
The division of labour and the handover are programmed and can be run directly.
One Agent finished reading the material and the results were handed over to the program. The procedure then transmits the results to the successor Agent or decides which branch of next step will follow.
Work that needs to be repeated can also be organized in a cycle.
An example of official documents is the revision of manuscripts: they are revised on an ongoing basis until the approval is made or the number of rounds is pre-established.
The principal, Agent, was assigned to read the report and then decided next step. Dynamic workflow, in turn, writes a large number of intermediate interfaces into the program and pushes them backstage.
During the waiting period, Major Agent can still communicate with users and see progress. Read the results after running, respond to the user or launch the next round of work.
However, after the return of the threads in workflow, the main Agent could not continue to follow up on the same threads as the caller for the regular sub-project Agent. It would therefore be preferable to schedule reconciliations and back-to-works ahead of schedule.
A sample of official contract review instructions provides that:
In the first round, each contract was awarded to one Agent; in the second round, the entire contract was re-examined by another Agent, not even those whose target terms were not found in the first round could not be avoided; the review was not adopted and was reviewed again on return.
This leaves an additional opportunity for discovery for the first round of missing clauses.
Of course, the contract cases show the way of work.
Anthropic did not disclose a specific division of labour in the Bug test posts in that group, on the basis of which it could not be concluded that the 66 Bugs had been identified through the same process.
A maximum of 1,000 Agents a single time is one of the most impressive figures for this survey.
It refers to a total of up to 1,000 Agents that have been activated by workflow over the life of the operation.
The current document lists simultaneous lines of work with a maximum of 64, and officially indicates that this number may be adjusted.
workflow can be carried out in successive groups, or can wait for the results of the previous phase to be returned and the latter scheduled.
Each Agent has an independent history of dialogue, sharing the documents and sandboxes in the sessions, i.e. the working environment for running codes and processing materials.
The developer can pre-configure the special Agent or make workflow a temporary definition according to the needs of the mission.
At the time of access, set type to multiagent 20261001, reconfiguring models, tools and mission requirements in the multiagent configuration.
When dynamic workflow is enabled, Claude is given the task to prepare a division of labour plan. The figure illustrates the screening of the change of control clauses in 300 contracts.
However, the division of labour was not sufficient and the unfinished business needed to be clarified.
An official example of a directive requires that a contract be classified as "uncovered" by a person who cannot read it, and that the rest continue.
"No problem identified" and "not found at all" must be separated. Otherwise, a coherent summary could easily hide the gaps in the mandate.
The developer can view the records of the operational phase and the various threads, and follow the log location problems.
In the case of contract reviews, the developers can track reading, reconciliation, etc., and view the execution records of each mission line.
Agent has more to do with sequencing and clarity as to which tasks have been completed and which gaps remain.
A distinction needs to be made: a maximum of 1,000 Agents were activated and 66 Bugs were identified on three occasions, referring to Managed Argentinas Dynamic workflow, not Projects.
Back to the initial tests: Agent found 14, 15, 27 Bugs three times, and Dynamic workflow 66 times.
Much more has been identified, but it is unclear how much more time and cost has been spent.
The model used, the actual number of Agents, time-consuming, Token usage and misstatement were not officially disclosed, and it is not yet possible to determine whether the workflow package is faster and more cost-effective.
When actually used, the costs of the two functions should also be viewed separately.
Projects uses the existing set of quotas, and both the work-line and the coordination dialogue consumes quantities. The more parallel tasks, the faster quota will be used.
Users can adjust the strength of models and reasoning or reduce Claude ' s lead. As a rule, the course of work will be suspended after the ceiling on the food package has been reached, pending the replacement of quota.
Managed Agens charges according to the model Token. In addition, an additional running fee of $0.08 per hour is charged for each session, calculated only for the time when the session is operational.
This $0.08 is only a running fee, and token, which consumes more Agent reading material, generates answers.
Developer can set the session budget and workflow model consumption is included. Operation was suspended when the budget was reached, but the model requests that had been made would still be completed and the final cost might exceed the budget.
The session budget covers model consumption in dynamic workflow and facilitates control of multiAgent collaboration expenditures.
Therefore, the official proposal begins with a clear mandate, mapping consumption and then gradually increasing complexity.
Agent's too much. It's not automatic.
How many real problems are identified, how many model transfer costs are added, and how much time it takes to review and correct, must be accounted for.
Less time is spent on work-sharing, and if more winding-up work were to be done, the project would not necessarily end sooner.
These two updates allowed Claude to start taking on more organizational work. Users report targets, budgets and acceptance criteria, and Claude is responsible for the division of labour, advancement and handover.
The next test is: can Claude give back the results that he can afford to examine after he has divided the work?
References:
https://x.com/ClaudeDevs/status/2108591328732856655
https://x.com/ClaudeDevs/status/2108591330129523146
https://x.com/ClaudeDevs/status/2108591331660468538
Editor: Won-woo Moses
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.