Complete AI Training
Start for Free

Your job's AI kit

Get your AI kit

Tell us who you are and what you do. We show you your kit right away and email you the link: skills, prompts, AI agents, MCP servers and courses for your job.

500+ jobs ready, and we make a kit for any other job. No payment needed to look.

Share

AI news ·

AWS DevOps Agent adds sandbox code execution, memory dreaming, and organizations support

AWS DevOps Agent now supports AWS Organizations, letting teams connect an entire organizational unit to an Agent Space instead of onboarding accounts individually.

AWS DevOps Agent reached AWS GovCloud (US-West) in November and added AWS Organizations support, letting teams connect an entire organizational unit to an Agent Space instead of onboarding accounts one at a time. The update also cuts investigation times, introduces sandboxed code execution, and adds quality measurement tools that give operations and development leads concrete ways to validate the agent's output before relying on it in production.

Faster investigations with real-time visibility

Investigations now complete faster and surface findings as they happen. Each investigation's Summary tab updates in real time as the agent collects evidence, showing initial findings, the current cause hypothesis, and next steps before the final root cause and mitigation plan appear. Teams using the built-in Slack integration receive these summaries directly in the relevant channel.

The new AWS DevOps Agent Sandbox, now in preview, gives the agent an isolated Linux environment where it can write and run code during investigations and chat conversations. The agent can analyze large log files, calculate statistics across metrics, or correlate data from multiple sources. Each sandbox is scoped to a single investigation or conversation and discarded when it ends. AWS calls from the sandbox are limited to read-only operations, and by default the sandbox reaches only AWS service endpoints until hosts are added to its allowlist. During preview, Sandbox is available in US East (N. Virginia), US West (Oregon), Asia Pacific (Tokyo), and Europe (Ireland).

Nine investigation metrics now publish to Amazon CloudWatch in the AWS/AIDevOps namespace with no configuration required. The metrics count investigations by end status, count completed investigations that found a cause or have a mitigation plan, record time to first cause and first mitigation plan, and record each Review result. Teams can create CloudWatch alarms on these metrics.

Measuring and improving investigation quality

Three launches help teams assess how well the agent performs and validate configuration changes. Completed investigations can now be replayed after changing skills, instructions, or integrations, then compared with the original by grader score, time to root cause, tools and skills used, and investigation timeline. Replays take no mutative actions and do not alter the original investigation. Because replays query the live environment, data available during the original investigation might not be available to the replay.

Reviews provide a quality assessment of completed investigations, returning a helpful or not helpful verdict and a pass, fail, or not applicable result for five dimensions: scope alignment, evidence grounding, completeness, environment impact, and accuracy. The agent now runs Reviews automatically on a sample of completed investigations - when a linked ServiceNow or PagerDuty ticket is resolved, seven days after an investigation completes, or both. Automatic Reviews are on by default with a 10 percent sampling rate, adjustable from 1 to 100 percent on the Grader agent's Configuration tab.

Code Change Risk Analysis, previously called release readiness review, is now in preview in all commercial AWS Regions where the agent is available. It reviews code changes for production risks before deployment by checking each change against your standards, its dependencies across repositories, its access controls, and your production environments. Each review produces a report with a recommended action, risk findings with affected code locations, and recommendations. Reviews can run in Kiro and Claude Code, automatically on GitHub and GitLab pull requests and merge requests, and on demand through chat.

Memory management that stays accurate over time

Teams can now add their own knowledge to a custom memory store in three ways: upload Markdown files already in the memory format, import runbooks and postmortems through chat, or keep a store in sync with a source by running a custom agent. When importing through chat, the agent proposes memories and a store for each, saving nothing until confirmed.

Search on the Memories tab now matches by meaning as well as exact text, with text matches listed first. Agents use the same search to find memories during their work. A new knowledge graph links related memories - such as an alarm, the component that owns it, and a reflection on a past mistake with that alarm - and agents follow these links to find connected context. A weekly task called memory dreaming reviews how agents used each claim in managed memories during recent investigations and custom agent runs. It deletes claims that no longer hold, merges duplicates, and resolves contradictions. Every change records its reason and evidence and creates a new memory version for review.

Extending the agent with connectors and published skills

Partner-managed connectors now connect AWS DevOps Agent to an MCP server that a partner hosts. The connector defines the server's endpoints and authentication methods, so teams choose an endpoint and enter only their credentials. Some connectors create an alert webhook that starts investigations from the partner's alerts. The partner operates the MCP server, and requests to it can leave the AWS Region of your Agent Space.

The Skills Catalog in the web app lists published skills - such as skills that investigate AWS Health events or check service quotas - and lets teams add a skill to their Agent Space in one step without a GitHub account or zip file. Catalog skills do not update automatically; teams can sync them to the current version. Amazon EC2 Image Builder also now integrates with the agent in the Image Builder console, letting teams troubleshoot up to 50 failed image pipelines at once.

Collaboration where teams already work

Chat now extends to MCP clients such as Kiro and Claude Code, as well as Slack and Microsoft Teams channels. Directed chat actions support GitHub, GitLab, Azure DevOps, ServiceNow, and PagerDuty MCP tools, and chat can use a sandbox to run scripts. Slack code channels give each investigation its own channel - teams and the agent work together without mentioning the agent in each message. In Microsoft Teams, users can now mention the agent in a private or shared channel to ask questions, explore AWS resources, and steer investigations, with the agent replying in a thread.

Why this matters for IT and development leaders

The combination of real-time investigation summaries, sandboxed code execution, and automatic quality Reviews changes how operations teams can validate AI-driven incident response. Instead of waiting for a final report, teams see the agent's hypothesis form in real time and can intervene earlier. Replays let platform engineers test skill and instruction changes against real past incidents before trusting them on live alerts - a capability that directly reduces the risk of configuration changes in production environments. For teams building their own diagnostic tooling, the sandbox means custom scripts can run inside the agent's workflow without building separate execution infrastructure. These features shift the agent from a black-box responder into a tool that teams can measure, tune, and extend with their own institutional knowledge.

Share