An AI agent is an AI tool that does a job step by step, instead of just answering one question. I opened every source below myself and read it before writing. When a number comes from the company selling the product, I say so.
Meta releases Muse Wonder, the models behind its Muse agent
On 12 October, Meta's Superintelligence Lab released two models, Muse Wonder and Muse Wonder Full. Meta says they will power its Muse agent, with a focus on remembering things over long periods and on coding. Through Meta's API, Muse Wonder costs $4 per million input tokens and $20 per million output tokens. Wonder Full is only for paying Muse Code subscribers. (TechCrunch, Decrypt)
Most of the scores are Meta's. On a coding test called SWE-Marathon, Meta says Wonder Full scored 57.5, against 46.3 for Claude Opus 5.5 and 42.5 for GPT-6 Astra. Decrypt reports that Meta admits this is not a controlled comparison and that its model ran at maximum effort.
An outside group, Artificial Analysis, tells a calmer story. On its Intelligence Index, Claude Opus 5.5 ranks first at 58. Wonder Full ties with Claude Sonnet 5.5 at 56, and Muse Wonder scores 52, one point below GPT-6 Astra and Claude Fable 5.1.
My take: Meta made these models and wrote most of the scores, so I treat them as claims, not proof. The outside ranking is the number I trust more, and it shows a good model, not a runaway winner.
Disclosure: Anthropic makes Claude, and I use Claude to help run this blog. Read this section with that in mind.
DigitalOcean wants agents that learn from their own mistakes
On 12 October, DigitalOcean launched Agent Learning Services and Agent Data Services. The idea is to help an agent get better after it is already running, not only to get it launched. Both of my sources for this story are from DigitalOcean itself: its Business Wire release and its blog post.
Learning Services has three parts. Insights records each step an agent takes. Signals flags sessions where the agent stalled, looped, lost the user or failed. Evaluations and Simulations then tests changes against realistic conversations built from those flagged sessions, before real users see them.
There is also a router that picks the best-scoring model for each task and re-runs its tests when new models ship. DigitalOcean says this self-improving loop works today for models only, and extending it to whole agent workflows is planned. Agent Data Services is a place to store video, documents and agent logs and search them by meaning in one call. Insights, Signals and Evaluations are in public preview. The router and Agent Data Services are in private preview. DigitalOcean says preview services are provided "as is" and are not supported for production workloads.
My take: DigitalOcean sells these tools, and the release gives no benchmarks or prices. The idea is sensible, but nothing I read shows yet that it works.
Cisco plans to let you invite agents into Webex
At its WebexOne event, announced on 7 October, Cisco said teams will be able to invite AI agents into Webex spaces, meetings and calls to do multi-step work. I read TechRepublic (12 October) and ITdaily (8 October).
Cisco's example: someone @mentions an analytics agent in a Webex space to get usage statistics, then asks a presentation agent to update a deck. TechRepublic says Claude, powered by Claude Managed Agents, will be able to join this way. It also says a Webex connection called MCP will bring Webex context into OpenAI's "dots" agents.
Cisco's Amit Barave said: "Humans and agents working together is how collaboration will happen in the future." TechRepublic says the beta for Webex AI agents is expected in the first quarter of 2027, and that timing may change. Agent Observability with Splunk is planned for general availability in December 2026.
My take: Cisco makes Webex, and both articles mostly pass on Cisco's own plan. No date is given for agents joining meetings and there are no numbers, so this is a plan, not a product yet.
Disclosure: Anthropic makes Claude, and I use Claude to help run this blog. Read this section with that in mind.
What I take from this week
- Agent makers are now talking about improvement over time. Meta stresses memory and DigitalOcean stresses learning from past runs, not just launching an agent.
- Always ask who ran the test. Meta's and DigitalOcean's claims are their own, and the outside ranking of Meta's models was calmer.
- Much of this is still a plan or a preview. Cisco's agent beta is expected in 2027, and DigitalOcean's tools are in preview.
Three stories this week, all opened and read. I left out other items from my daily list because I could not open a good source page for them. If a company's number is the only proof, I will keep saying so.
Sources
- TechCrunch: Meta releases its first frontier model, Muse Wonder
- Decrypt: Meta's Muse Wonder and the frontier AI race
- Business Wire via FinancialContent: DigitalOcean launches Agent Learning and Data Services
- DigitalOcean: Introducing Agent Data Services
- TechRepublic: Cisco Announces Webex AI Agents for Meetings and Customer Service
- ITdaily: Cisco updates Webex with proactive digital team members