A CRM is not only for you. It can also be the place where your AI agent tracks its own work and keeps a logbook of every change. Agents like Claude start each chat mostly fresh, so long-running jobs lose their thread. Give the agent a pipeline with stages such as Backlog, In progress, Waiting on human and Done, and it can log what it did, record why it decided something, write down the next step, and pick up tomorrow where it stopped. You watch progress on the board, and you can read the full story behind any card.
Below: why a pipeline works as both a task tracker and a library of changes, how to set up an "agent workspace" in the KMS AI CRM with a standing instruction for a Claude Project, and the honest limits. If you want the bigger picture of a CRM that you and your AI both work in, read AI CRM that works both ways.
The problem: agents forget between chats
Ask Claude or ChatGPT to help with a job that takes more than one sitting, like reaching out to 40 prospects or moving a website to a new platform, and you hit the same wall. A new chat starts without the details of the last one. Even with memory features in your AI app, a chat history is not a structured record of what is done, what is blocked, what was decided and what comes next.
So people paste long summaries into every new chat, or keep a notes doc that drifts out of date. You cannot see progress at a glance, and nobody can tell later why a choice was made.
The idea: a pipeline as the agent's task tracker
A CRM pipeline already has everything an agent needs to stay organized:
- Records for each unit of work (a prospect, a page to migrate, a ticket, a research question).
- Stages that show status.
- Fields for owner, next step and due date.
- A journal on each record: notes, calls and stage changes, each with a time.
- Reminder timers that flag work that has gone quiet.
Connected through MCP (Model Context Protocol, an open standard for linking AI apps to tools), the agent reads and writes all of this itself. Tomorrow, a fresh chat reads the same records and continues.
Meanwhile you open crm.kissmyskills.com and see the same board: in progress, waiting on you, done.
A library of changes: the agent's logbook
Tracking status is half the job. The other half is memory of how the work got where it is. The CRM keeps that in two layers.
Layer 1: each card's journal
Every record has a journal. Notes, logged calls and every stage change land there with a timestamp. When the agent adds a note after each step, a project card turns into the full story of that piece of work: what was tried, what failed, what was decided and why, and what comes next. Open the card for "Pricing page migration" and you can read it top to bottom like a short logbook, without asking the agent to remember anything.
Because stage changes are logged too, you also see the timeline: when the card entered In progress and how long it sat in Waiting on human.
Layer 2: the setup history
The board itself also has a history. Every change to the setup (stages, custom fields, reminder timers, a template you applied) is saved as a numbered version with when it happened, who made it and what changed. Version 0 is the original default setup.
This matters for an agent in three ways:
- It can read how the process evolved. Ask "show me the setup history" and the agent sees, for example, that a "Needs QA" stage was added in version 4 and the review timer went from 5 days to 3 in version 6.
- Rollback is safe. Rolling back restores an earlier version as a new version. Nothing is deleted, so you can always roll forward again.
- Destructive changes need you. Removing a stage or a field, or rolling back, is described first and only happens after you confirm.
Together, the two layers give you a library of changes: the journal tells you what happened to the work, the setup history tells you what happened to the process.
Decision records the agent can find later
The most useful habit you can give an agent is to log decisions in a fixed format. Ask it to write every real decision as a note like this:
Decision: Keep the old /blog URLs and add 301 redirects to /blogs/news.
Why: 140 posts have backlinks; changing slugs would lose them.
Alternatives: Rename all slugs (rejected, link loss); keep both paths (rejected, duplicate content).
Next step: Write the redirect list for the first 50 posts.Next week, when you, the same agent, a different agent or a teammate asks "why did we keep the old URLs?", the answer is one search away. The CRM searches records by text, including client names and notes, and can filter by stage. A search for "Decision: redirects" or "301" brings back the card and the note that explains it.
That is the difference between a pile of chat transcripts and a library: the decision sits on the work it belongs to, with a date, readable by anyone on the board.
How the agent resumes work
A good routine turns a fresh chat into a continuation. Here is the one we recommend.
Start of session
- Check today. The today view lists overdue next steps, cards with amber or red reminders, and upcoming dated items. This is the agent's to-do list.
- Search open cards. Look at In progress and Waiting on human to see what is active and what you may have answered.
- Read the latest journal entries on the card it is about to work on, including the last handoff note and any decision records.
End of session
- Write a handoff note: what was done, what is half done, what to do first next time.
- Update the Next step and Due date fields.
- Move the card to the right stage.
Reminder timers catch what slips. You can set a card to turn amber and then red after a number of days in a stage, or after a number of days since its last update. If the agent left a card stuck in In progress, it shows up red in the today view at the start of the next session, and on your board too.
Use cases
- Outreach. One card per prospect. The agent drafts messages for your review, logs each step, and moves a card to Waiting on human when a reply needs your judgment. The CRM does not send email, so you send and tell the agent what happened.
- Website migration or content plan. One card per page or article. The "Next step" field tells a new chat that the pricing page still needs its meta description, and decision records explain why the URL structure looks the way it does.
- IT tickets. One card per issue, with troubleshooting notes in the journal. When the same problem returns in two months, a text search finds the earlier fix.
- Research. One card per question, findings in notes, Done when the answer is ready for you.
How to set up an agent workspace
You can build this entirely from chat once the KissMySkills connector is added (see the connector page). Here is a layout that works for most agent jobs.
Stages
- Backlog: work that is known but not started. You or the agent add cards here.
- In progress: the agent is actively working on it.
- Waiting on human: blocked on a decision, approval or access only you have. The agent moves cards here.
- Done: finished and reviewed. You move cards here after review.
Fields
- Owner: agent or human.
- Next step: one short sentence describing the very next action.
- Due date: when the next step should happen.
Prompts to build it
Paste these into Claude or ChatGPT one at a time, or all together:
Show me my current CRM setup.Rename my stages to: Backlog, In progress, Waiting on human, Done. Show me what will change before removing any extra stages.Add custom fields: "Owner" (agent or human), "Next step" and "Due date."Set a reminder timer: amber after 2 days without an update, red after 4.Show me the setup history.
Each of these changes becomes a new version in the setup history, so if the layout does not suit you, ask the agent to roll back to an earlier version.
A standing instruction for your Claude Project
Claude Projects let you add instructions that apply to every chat in the project. Paste something like this into the project instructions, then edit it for your job:
You track your work in my KissMySkills CRM and use it as your logbook.
Start of every chat:
1. Check today: overdue steps, amber or red reminders, upcoming dated items.
2. Search cards in "In progress" and "Waiting on human".
3. Before working on a card, read its latest journal entries and handoff note.
While working:
- Work on one card at a time.
- After each meaningful step, add a short note: what you did and the result.
- Log every real decision as a note in this format:
Decision: ...
Why: ...
Alternatives: ...
Next step: ...
- Before deciding something, search the CRM for earlier decisions on the topic.
End of every chat:
- Write a handoff note: done, half done, what to do first next time.
- Update "Next step" and "Due date", then move the card to the right stage.
Rules:
- If you need a decision, approval or access from me, move the card to
"Waiting on human" and say exactly what you need.
- Never move a card to "Done" yourself; tell me it is ready for review.
- Create new cards in "Backlog" with Owner set to agent unless I say otherwise.
- Before removing a stage or field, or rolling back the setup, describe the
change and wait for my confirmation.The same idea works in ChatGPT, which also supports the connector. Put the instruction wherever your app keeps project or custom instructions.
A week with the agent workspace
On Monday you ask the agent to plan eight articles. It creates eight cards in Backlog with a next step on each, and logs a decision record about the topic order. You approve the list. On Tuesday you open a new chat and say "continue." The agent checks today, reads the board, picks the first card, drafts an outline, adds a note, and moves it to Waiting on human with the next step "approve outline." You check the board at lunch, see three outlines waiting, and approve two. On Wednesday the agent starts drafting those two. On Friday your editor Emma wonders why article four targets a different keyword; she searches the CRM and finds the decision note.
For an overview, ask: Summarize all cards by stage and anything overdue. The agent answers from analytics and the today view.
Honest limits
- The agent only updates when it runs. It does not work in the background by itself. It acts when you open a chat and ask, or when a scheduled task in your AI app starts a session.
- A human should review. Agents can skip a step, misread a note or mark something too optimistically. That is why the instruction above keeps "Done" in your hands.
- The logbook is only as good as the notes. The journal records stage changes on its own, but the reasons only exist if the agent writes them. The decision format makes that habit easy to check.
- It is not an orchestration framework. No built-in multi-agent handoffs, retries or workflow graphs. It is a shared, structured place to track work.
- No email sync or marketing automation. The agent logs what happened in the chat; it does not watch your inbox.
Pricing
The CRM is free for up to 300 records and 1 user, enough for most agent workspaces. The same connector also loads KissMySkills specialists, so you can say "load the sales specialist and work through the outreach cards." All-Access is $9 a month for unlimited records plus more than 2,500 AI skills, prompts and agents. Team is $29 a month for 5 seats, useful when several people review the same agent's work. For a wider comparison, see our guide to AI CRMs.
FAQ
How is this different from a notes document or a vector memory?
A notes doc is free text the agent must reread and interpret every time. The CRM gives each piece of work its own card, a status, a next step and a timestamped journal, visible to humans on a board, with a versioned setup. It is not a semantic memory store: it does not embed your notes or guess what is relevant; the agent finds things by text search and by stage. For fuzzy recall of past chats, use your AI app's memory. For a shared, checkable record of work and decisions, use the CRM. Many people use both.
Can I see exactly what the agent changed?
Yes. On each card, the journal shows every note, call and stage change with its time. For the board itself, the setup history lists every change to stages, fields, timers and templates as a numbered version with when, who and what. You can roll back to any earlier version; the rollback is saved as a new version, so nothing is lost.
Will the agent keep working while I am away?
Not by itself. The agent acts when it runs: when you send a message, or when a scheduled task in your AI app starts a session. When it runs again, it checks today, reads the board and continues from the last handoff note.
Does this work with ChatGPT and Claude Code?
Yes. The KissMySkills connector works in Claude (claude.ai, the desktop app and Claude Code) and in ChatGPT. In Claude Code it helps track a long technical job across sessions.
Want to try it? Create a free CRM, add the connector, paste the setup prompts above into a new chat, and give your agent the standing instruction.
