|
Hey there 👋
I opened two primaries before coffee. AWS’s own Whats New page on Bedrock Managed Agents, and Anthropic’s Claude Code changelog for 2.1.290 and 2.1.291. Neither is a keynote clip. Both are the kind of page you grep your stack against.
My bias up front: the interesting agent story this month is not a new demo. It is where the loop is allowed to run, and who owns the identity when it does. Today that story is AWS hosting an OpenAI-shaped agent runtime inside the permissions model you already fight with security about.
- Ron
The Big Thing
Amazon puts OpenAI agent loops behind your IAM
AWS and OpenAI shipped a public preview of Amazon Bedrock Managed Agents powered by OpenAI. The primary is blunt. BMA is built on a customized version of OpenAI’s Agents API, engineered to be AWS-native and integrated with AWS resources. You build agents optimized for OpenAI models that run entirely inside AWS with the identities, permissions, and governance controls you already use.
That sentence is the product. For a year, a lot of teams wanted OpenAI agent behavior and failed the internal review that asks where the agent runs, what role it assumes, and who can revoke it. BMA answers with the boring machinery: same account boundary, same IAM story, same Bedrock docs path.
During preview there is no additional charge for BMA beyond the underlying AWS resources your agents consume. Pricing can change at general availability. Preview regions are US East (N. Virginia), US West (Oregon), and US East (Ohio). AWS’s weekly roundup also notes execution choices: self-hosted compute you already run, or Amazon Bedrock AgentCore Runtime for managed runtime sessions and configurable storage in your account.
What this is not: a free OpenAI bill, a promise that every Agents API feature maps 1:1, or a reason to skip a threat model. Treat the joint engineering claim as a platform claim. Run one representative agent path in a non-prod account, map tools to least-privilege roles, and write down what breaks when you revoke the role.
Ship it? DEPLOY in non-prod this week if OpenAI agent loops were blocked only by “it lives outside our AWS boundary.” Skip if you are happy on a different runtime and do not need the OpenAI shape. WATCH the GA price sheet either way.
Sources: AWS Whats New, AWS News Blog, Bedrock docs.
Tour de Headlines
⏳ October 9: free Gemini drops to Flash-Lite only
Update: three days left. Personal accounts with no Google AI plan keep Flash-Lite and lose Flash and Pro from October 9. AI Plus ($4.99) keeps Flash-Lite and Flash but loses Pro; those users get an email with their personal switch date. AI Pro and Ultra keep the full set. If your free stack still assumes Pro or Flash for agent work, the week is short. Verdict: ACT if free Pro is load-bearing.
🔧 Claude Code 2.1.290 / 2.1.291: onboard Managed Agents, fix cloud permissions
Anthropic’s changelog is the primary. 2.1.290 (Oct 5) adds serverToolUses on a mod’s turn.step hook, agentId on tool.check, and /claude-api managed-agents-onboard for a URL or a Console quickstart name such as deep-researcher. 2.1.291 (Oct 6) fixes a regression where cloud sessions could drop answers to permission prompts. If you upgraded to 290 and cloud approvals felt flaky, pull 291 today. Verdict: DEPLOY if you live in Claude Code cloud or are wiring Managed Agents docs into the CLI.
📅 October 19: Copilot retires another model set
GitHub will deprecate, across Copilot Chat, inline edits, ask and agent modes, and completions: Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, and Grok 4.5 on October 19, 2026. Suggested replacements: Gemini 3.8 Flash, GPT-5.6 Sol, GPT-5.6 Luna, Grok 4.6. Business and Enterprise admins may need model policies flipped if default enablement was turned off. Verdict: ACT before the 19th if anything still pins the retiring set.
|
Sponsor
Your calls leave a signal. Read it.
RapportScore measures how your team communicates on real calls and coaches them on it. It scores what happened on the call, using human signals plus AI, so your people get better at the conversation itself, meeting after meeting.
See your team’s score →
|
Tool of the Day
🧭 HackerRank Chakra
An AI interviewer on a real-repo canvas. Humans still hire.
After about six months in beta, Chakra is generally available. HackerRank told TechCrunch the system ran more than 500,000 interviews in testing, with Snowflake, Snorkel, and Capgemini among early customers. CEO Vivek Ravisankar frames it as collapsing recruiter screen, take-home, and engineer interview into one sitting, and measuring judgment and “AI fluency,” not only correct answers. Suspicious-activity flags were 70% to 80% lower than traditional assessments in their comparison. That range is the vendor’s, and it varies by geography and seniority. Grade it like any internal metric until you see your own funnel. Legal and bias review is not optional: an agent that scores candidates is a different risk class than an agent that refactors a repo.
Read the TechCrunch report →
Worth a Click
The agent loop is getting portable. OpenAI’s shape now has an AWS address. Claude Code can onboard a Managed Agents pattern from a URL. None of that removes the hard part.
My bet: the teams that stay sane this quarter pick one identity boundary per agent workload and refuse a second “just for the demo” plane. Everyone else will debug two permission models while the clocks on free Gemini (Oct 9), GPT-5.5 product surfaces (Oct 14), and Copilot pins (Oct 19) keep ticking.
The demo can tour. Production needs a role you can revoke.
- Ron
| 
