Some Agentic AI Options To Explore Responsibly
Estimated Reading Time: 3 minutes
Last updated on 5th April 2026
Agentic AI is one of the faster-moving corners of the field right now, and also one where I am most inclined to proceed carefully. Where most AI tools wait to be asked something, agents are designed to act on their own initiative, chaining together tasks, reaching into external services and doing things without a prompt for every step. That is genuinely useful, but it is also the quality that makes them worth approaching with clear eyes. The stories that circulate about poorly contained deployments, unintended actions and data ending up where it should not are not just edge cases. They reflect something real about what happens when software is given broad permissions and told to get on with it.
Here is a word of caution before diving in: these tools can do a lot, and that is precisely why the defaults, the permissions and the containment matter so much from the outset. After all, Peter Steinberger himself acknowledged that he deliberately kept the installation process for OpenClaw non-trivial so that users would pause and understand the basics: what AI is, that it can make mistakes and what prompt injection means. That is good instinct when running an agent that touches email, calendars or messaging services is handing over meaningful access, and the time to think about what that means is before the agent is running, not after.
This open source autonomous AI agent framework by Peter Steinberger lets people run a persistent assistant on their own computer or server that carries out practical tasks rather than only chatting. It can link with messaging services such as WhatsApp Telegram and Discord plus email calendars and other tools to do things like handling inboxes booking meetings checking in for flights controlling smart home devices and running local commands, with configuration and state stored to retain preferences and continuity. Built around large language models and a skills approach that is modular in its nature, it can also work with local model servers such as Ollama for private use without cloud costs. After going viral on GitHub in early 2026 and as its creator joined OpenAI with plans for a move into an open-source foundation, it also drew scrutiny over security, since exposed or poorly isolated deployments and broadly permissioned extensions can leak sensitive data or be abused unless run with proper containment such as containers or virtual machines.
Designed as an autonomous AI agent, Manus aims to take a goal and then plan, execute and refine multistep work with limited supervision, using tools such as web browsing, coding, file generation and external services to produce more complete deliverables than a standard chat assistant. It is used for tasks like research and report writing, building small applications, analysing data, comparing products and automating routine digital workflows, although it remains early stage, can drift on longer runs and depends on clear instructions. Developed by Monica, it has become part of the broader shift towards agent-based systems and, by 2026, discussion has moved from initial excitement to practical integrations such as deeper Google Workspace editing alongside wider questions about ownership and control, following Meta completing an acquisition said to be worth about 2 to 3 billion dollars and subsequent regulatory scrutiny, particularly in China, around cross border technology transfer.