CTA keyword: AGENT · from the @compedge.ai x @am_roshh carousel
On 8 Oct 2026 Google Cloud announced the Gemini agent: one agent for work that takes objectives, runs sub-agents for hours or days, keeps four kinds of memory and picks Gemini or Claude models for each job. Agents like this only work as well as the brief you hand them. Below: a goal template, a memory checklist, the guardrails to set before it starts, and the questions to ask any vendor. It works whether you use Gemini agent, Claude, ChatGPT or something else.
What Google announced
- One agent for work. Google Cloud calls the Gemini agent a "single, universal agent for work" that answers questions, handles knowledge work, creates images and media, and writes and runs code. Its pitch: "You give it objectives, not instructions." (Google Cloud blog, Thomas Kurian, 8 Oct 2026)
- It runs for hours or days. Google says it can create temporary sub-agents for multi-step tasks, with parallel and sequential steps that "can run for hours or days", and keeps running in the cloud after you close your laptop. (Google Cloud)
- Four kinds of memory. Session memory for the task in front of it, semantic memory (a knowledge base it builds as it reads and talks to people), procedural memory for how a job gets done, including skills it writes for itself, and episodic memory of everything it has done before. (Google Cloud; 9to5Google)
- It lives where you work. Inside Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar, and reachable from the command line, Slack and Microsoft 365, or through an API as a "headless" agent. (Google Cloud; 9to5Google)
- It picks the model per job. Google says it runs each job on the model that fits best: the Gemini family and Claude models from Anthropic today, other private and open models later. (Google Cloud)
- Controls Google lists. Each agent gets its own identity, role-based permissions, an audit trail of every action, a sandbox, and per-project spend caps that pause the agent when hit. (Google Cloud)
- Availability. Private preview now. 9to5Google reports wide availability "soon" for Workspace customers on select Business and Enterprise plans.
Everything from here down is CompEdge's own guidance, not Google's. It applies to any agent that works on its own for a long time.
1. The goal brief (copy, fill, paste)
An agent that runs for days cannot ask you what you meant at 2 am. Write the brief like you would for a new hire who starts tonight.
OBJECTIVE
[One sentence: the outcome,
not the steps. e.g. "A 6-slide
update deck on Project X for
my manager by Friday 5 pm."]
WHY IT MATTERS
[Who uses the result and for
what decision.]
DONE MEANS
- [Format: deck / doc / sheet]
- [Length or size limit]
- [Must include: A, B, C]
- [Must cite: where each
number came from]
INPUTS YOU MAY USE
- [Folder / doc / thread]
- [Data source]
NEVER
- [Send anything outside the
company without my yes]
- [Touch: list of folders,
clients, systems]
- [Spend more than: $___]
CHECK IN WITH ME
- [When you are 50% done]
- [Before any email, invite
or payment goes out]
- [If you are stuck > 2 hours]
DEADLINE
[Date, time, time zone]
Tip: if you cannot write the DONE MEANS block, the task is not ready to delegate. Do that part first.
2. Memory checklist: what to give it, what to check
Google names four memory types. Most long-running agents have some version of each, even if they call them something else. Use this list before a multi-day job, and again after it finishes.
SESSION (this task)
[ ] Brief pasted in full
[ ] Links to every input
[ ] Deadline + time zone
SEMANTIC (what it knows)
[ ] Company terms defined
(e.g. "active customer")
[ ] Old / wrong docs removed
from what it can read
[ ] One source of truth per
number
PROCEDURAL (how we do it)
[ ] Your template or a past
good example attached
[ ] Steps it must not skip
(review, sign-off)
[ ] Skills it wrote for
itself reviewed by a human
EPISODIC (what it did)
[ ] You can see its history
[ ] You know how to delete
a memory that is wrong
[ ] Past mistakes noted so
it does not repeat them
Why this matters: an agent that remembers a wrong fact will repeat it for weeks. Check what it learned, not only what it delivered.
3. Guardrails before you press go
Set these before the first long job, not after the first incident.
- Its own identity. It should act as itself, not as you, so its actions show up under its own name in logs.
- Least access. Give it the folders and tools this job needs, nothing more. Add access later if it asks.
- A spend cap. Put a hard limit on what one project can spend on tokens and compute. Agents that loop can burn budget overnight.
- A human yes for anything external. Emails to clients, calendar invites, payments, public posts: draft only, you send.
- An audit trail you have opened once. Find where its actions are logged before you need them.
- A kill switch. Know how to pause or stop it in under a minute.
- A first-week review. Read everything it produced in week one. Trust grows from evidence, not from the demo.
4. Vendor-risk checks (ask before you buy)
Use these with any agent vendor: Google, Anthropic, OpenAI, Microsoft or a startup. Write the answer and the date next to each one.
DATA
[ ] Which models see my data?
(one vendor or several?)
[ ] Is my data used to train
any of them?
[ ] Where is it stored, and
for how long?
ROUTING
[ ] Can I see which model ran
each job?
[ ] Can I pin or block a
model for some jobs?
CONTROL
[ ] Own identity per agent?
[ ] Role-based permissions?
[ ] Full audit log I can
export?
[ ] Spend caps per project?
EXIT
[ ] Can I export its memory,
skills and history?
[ ] What happens to my data
if I cancel?
STATUS
[ ] Preview or generally
available? (preview =
features and terms can
change)
Red flag: "we pick the best model for you" with no way to see or set which one. Model choice is a feature only if you can audit it.
5. Five jobs to hand an agent first
Start with work that is useful, easy to check and safe if it goes wrong.
- The weekly update. Pull status from your docs and threads into one page. Easy to check against what you already know.
- Meeting prep. Brief me on everyone in tomorrow's meetings and the open threads with each.
- Research to a table. Compare 5 vendors, tools or competitors on criteria you set, with a source link per cell.
- Inbox triage (draft only). Sort, summarize and draft replies. You press send.
- Clean up a messy doc or sheet. Same content, your template, with a list of everything it changed.
Leave for later: anything that sends money, signs, deletes, or talks to customers without you.
Your first-agent sheet: fill this in
AGENT / TOOL: ____________
STATUS: [ ] preview [ ] GA
FIRST JOB: ____________
BRIEF WRITTEN: [ ]
DONE MEANS CLEAR: [ ]
ACCESS LIMITED: [ ]
SPEND CAP SET: $____ [ ]
EXTERNAL = MY YES: [ ]
AUDIT LOG FOUND: [ ]
WEEK-1 REVIEW DATE: ________
Sources
- Google Cloud blog, Thomas Kurian, "Welcome to Gemini at Work 2026: Introducing the Gemini agent", 8 Oct 2026: https://cloud.google.com/blog/products/ai-machine-learning/welcome-to-gemini-at-work-2026
- 9to5Google, Abner Li, "Google Cloud announces 'Gemini agent' as 'universal agent for work'", 8 Oct 2026: https://9to5google.com/2026/10/08/gemini-agent-google-cloud/
The "What Google announced" section reflects Google's post and the 9to5Google report of 8 Oct 2026. The brief template, checklists, guardrails, vendor questions and job list are CompEdge's own.