AI Agents That Do Work: What Changed This Week
Meta Muse passed 600k downloads, Google CC became family helper, OpenAI shipped Agents API beta. What changed when agents started doing work, in plain words.
On this page
Chatbots answer, agents act. In one week three launches made that shift concrete: Meta Muse passed 600,000 downloads in five days, Google refocused CC on families on September 18, and OpenAI opened the Agents API beta on September 10. Here is what each one does, explained simply, and what to copy if you ship for users.
Takeaways
Muse proves demand for doers: email, bookings, forms, purchases, all from a chat thread with named agent and avatar. CC proves permissions are the product: each person picks what to share, CC gets its own account with scoped access. Agents API proves harness is heavy: compaction, tool search, parallel calls, subagents, sandbox choice, all managed. Startup lesson: sell the completed task, gate every side effect, price per completed run.
What can these agents actually do for you?
Muse books and buys: restaurant reservations, travel holds, form fills, plus Link by Stripe checkout, with Shop Pay planned. CC runs the household: permission slips, activity PDFs, supply lists, meal plans, drive times, shared Docs and Sheets. The Agents API runs the backend version: investigate a 5xx spike, split deployment plus error plus dependency checks across three subagents, save findings to /workspace/outputs.
old assistant: you ask -> it writes -> you copy -> you act
new agent: you ask -> it reads calendar plus email -> it acts -> you approve money moves
your job: decide what it may touch, what needs approval, what stays in group memoryPlain words glossary
Harness means the loop around the model: context packing, tool routing, retries, subagent fan out. Sandbox means the computer where code runs: OpenAI hosted, your VPC, or partners like Cloudflare and Modal. Compaction means summarizing old context so long runs fit.
How do permissions keep this safe enough to use?
Give the agent its own identity
CC gets its own Google account, Muse sits behind Sentinel: the agent acts as a guest with listed access, never as you with full access.
Share by sender, not by inbox
Forward school or club mail, or auto share named senders: CC suggests new ones weekly. Narrow sharing beats full mailbox sync for trust.
Gate money and outbound posts
Purchases, bookings, and form submissions pause for approval with missing detail prompts. Muse adds a $130,000 prompt injection bounty as backstop.
Keep group memory visible
CC updates shared memory after each task: show what was remembered, allow delete. People accept memory they can inspect.
When should a startup use the Agents API versus self host?
Use hosted when runs last hours, touch files and code, and need parallel subagents without building orchestration: one sessions.create call with model: gpt-6-astra, MCP tools, max_concurrent_subagents: 3, vault plus environment config. Self host when data must stay in your VPC, or cold starts and GPU shape dominate cost: the same harness talks to your compute through partner sandboxes. Either way you pay tokens plus tools, no extra API fee during beta, and the harness stays open source for inspection.
What is still missing?
Who cannot use these yet?
Muse is US only, 18 plus, with glasses marked coming soon. CC is US Gmail only, 18 plus, waitlist for new users, so teens on school accounts sit out even though classrooms run on Google.
Do I need subagents on day one?
No. Start with one agent plus tool search and programmatic calls. Add multi_agent only when tasks split cleanly into parallel pieces with separate context, like the 5xx example above.
As of September 22, 2026: agents moved from answering to finishing, and trust moved to permissions plus memory controls. Pick one repeatable task, scope access, require approval for spend, and log every action. Next, read why coding agents got hacked before you connect those agents to code.