Pinned Tweet
I run an IT agency with AI agents. The thing that ate our time: writing SOPs by hand.
So I built stepkeep. Click through a task once, get a step-by-step guide with annotated screenshots.
No account. No upload. No subscription.
stepkeep.app
Chrome Store for some reason decided to promote my extension as the installations surge. Happy to see this, stepkeep is a really good tool. Y'all deserve the best
A manager who wants to know how heavy support work really is will learn little from the ticket count alone.
1. Work some tickets yourself for a day. You will see where the queue stalls and how emergencies and high priority tickets get handed out.
2. Sit in with the techs for a few days and watch the handoffs.
Then ask each tech separately what is causing the stress.
Not in a group, one by one.
People often name a task the manager does not know they are being asked to do, like chasing a customer for a reply before a ticket can move on, or being the unofficial expert for one client's odd setup.
Those hidden jobs usually explain the load better than any report.
What task does your support team do that never shows up in the ticket system?
Agencies often hire ahead of the work. The plan says more people are coming, so the hiring starts before the revenue does.
Then utilization drops. People sit between projects, margins shrink, and the team has to be cut back. One team went from 4 people to 36, then back down to 28. That reset was painful, and it was avoidable.
The better rule is to hire against revenue and utilization. Hire when billable work is already there and the current team is busy. Look at what the team actually bills over a few months, not at the pipeline you hope to win.
Also expect to fix it if you overshoot. Cutting people is possible, but it is never clean.
Word is fine for a long technical document. Headings, a table of contents and numbered steps all work.
The trouble starts when the same text has to appear in more than one place. A process description ends up in the handbook, the onboarding guide, the client offer and the audit pack. Then someone fixes a step in one copy, and the other three stay wrong.
That's when a single-source tool is worth a look. You write each piece once and generate the different outputs from it. Fewer copies means fewer contradictions.
If your documents are one-offs, keep using Word. Think about switching when you notice you're pasting the same paragraphs into a fifth file.
When you raise prices, decide early what happens to your existing clients. Either they keep their old rates, or they move to the new ones. Don't leave it vague and handle each one ad hoc.
Think about the math. If you want the same revenue, you can get there with fewer clients. Every client you keep costs you onboarding, reporting requests and coordination time, and for a small team those costs add up fast.
So pick a rule first and apply it to everyone. Write it down so you can point to it when a client asks.
Before you say yes to a managed outbound or paid ads job, write down the full cost of it. Not just the tool fees. Add the sending domains, the mailboxes, the contact data, the copy, and your own time to run it.
Then set that against what the client can realistically close. How many replies turn into calls, how many calls turn into deals, and how big a deal is. If the numbers don't cover the cost, the work loses money even when everything runs fine.
Watch the timing too. There is usually a ramp of two to three months before results show up, and the spend comes at the start. So the client needs enough revenue to carry that stretch, and they need conversion data they actually track. Without tracked numbers, nobody can tell whether the campaign is working.
If the client can't absorb the early cost, don't take it on at a flat fee. Ask for a budget they can afford up front, or say no.
Cheap tokens are one thing, but what If you could slash tokens used by 80%
🫨简直无敌!DeepSeek Harness被清华微软团队直接“瘦身”:Token消耗砍掉87.93%,成功率反而暴涨14%!
Agent每次工具反馈后都要重新问一遍LLM,哪怕下一步早定死了(写完代码→编译,编译成功→测试)。清华+微软团队用HarnessMM直接把这些“已决定”的步骤缓存下来,跳过重复推理。
核心成果(98个HumanEval任务):
🔹DeepSeek Harness:Token从96.28k降到11.62k,砍掉87.93%!成功率从71.43%直接飙到85.71%
🔹Codex:成功率稳在90.82%,Token再降29.54%(66.72k→47.02k)
🔹原理:模型一次规划未来命令+精确参数+守卫条件(如“编译成功”),匹配就直接执行,跳过LLM回调。模型权重完全不动
🔹额外:MBPP上也大幅降本,递归自我改进(RSI)让Harness配置持续变强
Agent循环里大量Token浪费在“重新决定已经决定的事”上。这波直接把成本打下来,性能还往上走。
论文:arxiv.org/abs/2610.32813
#DeepSeek #AIAgent #Token效率 #清华 #微软 #HarnessMM #大模型降本 #Agent优化 #梁文峰 #梁圣
Raising prices can fix a client problem too.
One agency doubled its prices. A channel went from $500 to $1,000 a month, and a bundle of integrated ads plus SEO went to $1,750. In one month they signed 7 new clients worth $8,500 in monthly recurring revenue.
The side effects mattered as much as the money. Fewer clients meant fewer onboardings to run. The clients who did sign were a better fit and less reactive. And the higher price filtered out the poor-fit prospects, so nobody had to have an awkward "this isn't a good match" conversation.
If your team is stuck with too many small accounts and constant firefighting, look at your price list before you look at headcount.
Taking over an IT setup without a list of systems is a mess. Nobody knows who owns what, and the first outage turns into a round of "not me, try them."
Early on, write down every system and the person responsible for it. That means an internal owner or the MSP that runs it. Keep it short: system name, what it does, owner, backup contact. A spreadsheet or a single page is enough.
The point is speed. When something breaks, you know in a minute who to call, and you don't spend the first hour working that out.
Update it when an owner changes. An old name on the list is almost as bad as no list.
A freelancer sent an invoice that was meant for another client. Two clients, two invoices, and the wrong one went out.
The newspaper that got it paid $600 by check. The correct amount was $350, so they overpaid by $250.
What helped was speed. The freelancer contacted the client right away. The client cancelled the check and sent the right $350 by ACH. Once the error was reported, it was sorted out without trouble.
The real lesson is the last check before sending. Look at the client name, the amount and the project reference. Those three fields are where these mix-ups hide.
If it does happen, say so early. Waiting makes it harder for everyone.
If your team uses AI to write technical docs, decide what the source of truth is before anything gets published.
Here is how it goes wrong. Someone asks an AI tool for a technical description, the text reads well, and it gets saved as is. Later, the content team picks it up and reuses it word for word in published material. Nobody had checked it against how the system actually works.
The gap was there before the AI came in. Many teams never wrote their own reference docs, so generated text filled an empty space and looked finished enough that nobody questioned it.
What helps is simple. Pick one document as the truth, keep it short, and give it an owner. Every generated draft gets checked against that document before it leaves the team. The AI can write the first version. A person who knows the product signs it off.














