About 9 minutes to read.
What is up people!
Imagine asking AI to prepare a proposal for six short videos. You give one agent the client’s notes, another the scope, and a third the job of writing the final document. Each finishes its part. The proposal looks ready to send, until you notice that Friday’s deadline has changed meaning somewhere along the way.
The client wanted approved videos by Friday. The proposal promises first drafts by Friday. The summary passed between agents said “delivery Friday,” and the distinction disappeared.
You now have three completed tasks and a document you need to repair. Before adding another agent to catch the mistake, it is worth asking whether the work needed to be split there at all.
Anthropic’s October 22 guidance on simplifying agents makes that question timely. Its team describes cases where bringing specialist work back into one agent improved the result. The useful lesson is to examine what crosses the boundary between agents, then keep or remove that boundary according to the work it helps you finish.
📌 TL;DR
Keep connected decisions together. If writing the scope depends on the exact client conversation, preserve access to that conversation through the writing process.
Separate work that can stand on its own. A focused research question or a check against the original requirements can justify another agent. Giving it a job title does not establish that benefit.
Include your repair work in the result. A faster first answer is a poor improvement if you spend the saved time finding missing facts and stitching the work back together.
Keep the facts beside the decisions they change
A summary has to leave things out. That is why it is useful, and also why it can become a weak link when the next task depends on a detail its author did not recognize as important.
In the fictional video proposal, “delivery Friday” sounds like an adequate summary until someone must decide when editing starts, how much time the client gets to review, and whether revisions fit. The original phrase, “approved videos by Friday,” affects all three. A writer who sees only the shorter version can produce a coherent schedule around the wrong requirement.
This is a more specific problem than an AI being forgetful. The fact may never have reached the agent making the decision. Asking that agent to be more careful will not restore information it never received.
It helps to distinguish an agent from a set of instructions. A separate agent generally works with its own conversation and the information supplied to it. A skill gives an agent a procedure it can use while continuing the current work. The Agent Skills documentation describes loading the relevant instructions when needed, with supporting references or scripts available alongside them.
For the proposal, the procedure could explain how to turn a client request into deliverables, assumptions, a schedule, and open questions. The agent preparing the proposal can consult that procedure while still referring to the original notes. You gain a repeatable method without necessarily handing the job to another conversation.
That leaves a practical choice. Would a second agent bring something the first lacks, or would it mostly read a shortened version of the same material and continue the same decision? In the second case, keeping the work together is a sensible starting point.
You can also improve a handoff without removing it. Attach the original notes, identify the exact passage about timing, and ask the receiving agent to check the proposal against those notes. A concise handoff can point to evidence instead of attempting to replace it. Whether the agent can open the referenced material still needs checking; a filename in a message does not transfer the file.
The same reasoning applies when you manually move work between chat windows. If you keep copying your correction into three conversations, that is a useful clue about how much the tasks share. You may be maintaining an artificial separation that gives you more coordination work than help.
Give another agent a job it can finish independently
There are good reasons to divide work. Suppose the same proposal includes filming at a location whose published requirements you need to check. A separate research task can inspect that location’s official information and return the relevant restrictions, source links, and unanswered questions. It does not need to rewrite the client’s entire proposal to do that well.
Meanwhile, the main agent can work on the editing scope. Once the research comes back, it can incorporate the verified requirements into the same document. The two tasks have different immediate inputs and a clear point at which their results meet.
Compare that with asking a “strategy agent” to decide what the client needs, a “scope agent” to translate the strategy, and a “writing agent” to express the scope. Those labels sound organized. But if each needs the full conversation to interpret the job, the labels have not made the underlying decisions independent.
Earlier research provides a useful check on both enthusiastic delegation and enthusiastic simplification. In the April revision of Towards a Science of Scaling Agent Systems, researchers from Google, MIT, and other affiliations compared 260 configurations across six task benchmarks. They varied how agents coordinated while controlling aspects such as tools and prompts. Multiple agents helped on some work that could be divided into separate parts, while sequential planning performed worse under the tested multi-agent arrangements.
These were benchmark tasks, not a trial of your business. The authors also note limits in the tested designs and that prompts tailored to particular architectures might change the results. The evidence supports choosing according to the task. It does not establish an ideal number of agents for everyone.
A separate reviewer can be useful for a different reason: its job is to challenge the finished proposal against the original request. Give it both documents and ask it to identify unsupported promises, missing requirements, and contradictions. It can quote the relevant passages so you can inspect each finding.
You do not need to provide every explanation the drafting agent offered for its choices. The reviewer needs enough context to judge correctness, including the client’s constraints, but its task is to examine the document rather than defend the process that produced it.
An approving review still needs substance. “Looks good” tells you little about whether Friday means a draft or an approved video. A useful finding identifies the two conflicting statements, explains their consequence, and leaves you with something you can check yourself. Agreement between agents is helpful only to the extent that the supporting evidence holds up.
For a small proposal, a careful human read may be easier than arranging another AI review. For a long document with requirements scattered across several files, a separate pass may earn its cost. Choose based on what makes a mistake easier to find.
Count the work you still have to do
Simplifying can break a workflow too. Anthropic’s October 22 article describes a sample support agent that became faster after a change, but then answered an order question without asking the shipping specialist for the delivery date. A step that had been working stopped working, and the failure appeared in one of two runs.
That example is valuable because it prevents an easy conclusion: deleting an agent is not automatically an improvement. If the removed agent was the only place a necessary check happened, the check must survive somewhere else.
Return to the video proposal. You could combine its drafting stages and get a shorter process, yet still lose the deadline if the main agent does not read the original brief. You could preserve all three stages and fix the problem by passing the source material correctly. The thing to investigate is the missing information and responsibility, not the number on the agent counter.
Start with the finished document. Does it describe six videos? Does it distinguish first drafts from approved versions? Does it leave room for client review? Does it flag terms the client has not supplied instead of inventing them? These questions are much closer to the real job than whether every agent marked its task complete.
Then count the work left for you. Perhaps one version takes longer to produce but needs one minor edit. Another arrives quickly and sends you back through the notes to reconstruct the schedule. Recording those differences gives you a better basis for choosing than the speed of the first response.
Keep the errors visible rather than collapsing everything into a single score. A missing comma and an unsupported delivery commitment both count as corrections, but they have different consequences. For this proposal, preserving the commitment belongs ahead of making the prose prettier.
If the main agent keeps missing a requirement despite having the right source, adding a reviewer may be justified. If the reviewer keeps finding the same omission, fix the drafting instructions or source access as well. Otherwise you have built a reliable way to discover a recurring mistake while continuing to produce it.
Also this week...
Anthropic donated the Agent Skills standard to the Agentic AI Foundation on October 22. The foundation sits under the Linux Foundation, and the announcement includes a proposed roadmap for versioned releases, a formal proposal process, and security guidance.
For someone building useful procedures, the appeal is continuity across compatible tools. The instructions for preparing a good proposal need not be reinvented whenever you change the agent that reads them. That does not mean its app connections, permissions, or execution environment will transfer automatically. Preserve the procedure, then check what the next tool can actually do with it.
💡 Try this with one finished job
This is a suggested rehearsal, not a workflow tested for this issue. Use a harmless sample or a copy of a completed job whose requirements you already understand. Keep it away from live sending, publishing, payments, or customer records.
Save the brief and the checks before starting. For the fictional proposal, state that there are six videos, source footage arrives Monday, the client wants approved versions by Friday, and subtitles are required. State that the number of revision rounds and rights for paid advertising have not been agreed. A satisfactory proposal preserves the known facts, raises those two open questions, and identifies the need for a review schedule without inventing one the client has accepted.
Run your usual arrangement once. Use the same brief and ask for a proposal draft. If your usual method uses separate research, scope, and writing conversations, keep that structure. Save what each stage receives and returns. Do not correct missing facts during this first attempt; you want to see what reaches the end.
Run one conversation on the same materials. Start fresh, provide the identical brief and instructions, and ask one agent to prepare the proposal without delegating its drafting stages. Keep the model, available tools, and settings the same where your application allows it. Give neither version extra examples or feedback before saving its first result. If the application chooses models or agents for you, record that difference rather than treating this as a clean comparison of agent count alone.
Check both drafts against the saved requirements. Inspect correctness first, then note how long the work took and what you would need to repair. Because you are seeing the same job twice, familiarity can make the second review feel easier. Use the written checks to compare omissions, and treat timing as a rough observation. One attempt also leaves room for model variability. Repeat on another representative job before reorganizing a workflow you depend on.
Here is a starting request for the single-conversation version:
Prepare a proposal draft from the attached client brief. Keep the drafting work in this conversation and use the original brief throughout. Distinguish confirmed requirements from suggestions and unanswered questions. Do not invent agreement on deadlines, revision rounds, or usage rights. Before the final draft, compare it against the supplied requirements and flag any conflict you cannot resolve from the source. Produce a draft only; do not send messages or change external records.
The task and quality requirements should also appear in your usual arrangement. The single-conversation instruction is the intended difference. Adding stricter checks only to that version would test better instructions as well as a different division of work.
If you do not already have a multi-agent workflow, there is no need to build one for this exercise. Start with the single conversation and inspect its result. Introduce a separate task when you can identify a concrete limitation it would address, such as a substantial research branch or a document check you want performed independently.
🧰 Tools to try
Anthropic’s single-agent-loop kit, for readers who already build agents. The public repository includes a sample support system, comparison scripts, and recorded example runs. Read the example outcomes before considering changes to your own setup. Its free offline check verifies plumbing, not model quality; real evaluations can incur model costs.
The repository is an unmaintained reference implementation. It documents only two real-model example runs by one person, and says the finished simplified version has not been tested on a real model. That makes it material to inspect and adapt carefully, not a ready-made migration to trust with production work.
The official Agent Skills guide, for one procedure you already repeat. Start with something narrow, such as the proposal requirements above. A useful first procedure describes the needed inputs, how to handle unknowns, and what a finished draft must contain. Run it on an ordinary case before collecting more skills. Another folder is easy to create; a procedure that consistently saves you corrections takes evidence.
When you next open that six-video proposal, begin with the Friday deadline. The document should make clear what must be approved, what review time still needs agreeing, and which terms remain unanswered. Once those details are in the right place, you can decide whether the agent arrangement helped you get there or left you doing the connecting yourself.
P.S. An AI wrote these sentences from this week’s primary sources and earlier research. Payton set the editorial standards and what good looks like. He wants to be transparent when AI wrote the issue.