Stop Restarting Your AI Agent. Just Talk Over It.

Most people still talk to AI like it is a vending machine. Put in a request, wait, take whatever drops out, and if it is wrong, start over from the top. That model just broke.

The Vending Machine Model Of AI Is Over

Here is the part nobody warns you about when you hand a task to an AI agent: the moment you notice it going sideways, your only real option used to be to stop it and start again. You watch it draft the wrong email, order the wrong quantity, or head down the wrong research path, and you cannot say anything until it finishes. You just wait, then clean up.

On August 3, Nous Research shipped Hermes Agent v0.20.0, nicknamed "The Herald Release," and the headline change is exactly this problem. According to the release notes, the update ships across 3,650 commits and 1,400 pull requests from 647 contributors. That is not a patch. That is a new way of working with an agent.

The single most useful change in it has nothing to do with voice, even though voice is what got the most attention on social media this week. It is this: you can now type a correction into Hermes while it is mid-task, and it steers itself with your new guidance instead of stopping and restarting. Nous Research's own team calls it mid-turn redirection. Think of it less like undo, and more like tapping someone on the shoulder while they are still talking.

What Actually Shipped, In Plain English

Strip away the changelog language and five things matter for a business owner, not a developer:

Real interruptible voice. The old voice mode made you speak, wait, then listen to one long reply. The new one streams the reply clause by clause as it is generated, and you can talk over it mid-sentence. Say "Hey Hermes" and it wakes up hands-free, useful if you are walking a stockroom or driving between job sites and cannot look at a screen.

Mid-turn redirection. Covered above. This is the one to build around.

Agent-to-agent handoff (A2A v1.0). Before this release, if you wanted two AI agents to work together, you were the messenger, copying output from one and pasting it into the other. A2A gives agents a shared, standard protocol to discover each other and hand off work directly. One agent can now finish a task and pass the baton to a second, more specialised agent without you in the middle.

Signed outbound webhooks. The agent can now push a verified, tamper-proof notification (session started, task finished, a tool ran) to any endpoint you register, including something as simple as a Slack channel. That is your audit trail without you having to read every conversation.

Grounded citations. When Hermes researches something now, it can attach a verifiable source to each claim instead of producing a plausible-sounding answer with nothing behind it, closer to how Perplexity cites its sources.

Put the three that matter most together, correction, handoff, and a log of it all, and you get a pattern no single-agent setup could do cleanly before this week.

The Automation Worth Building: Two Agents, One Correction Window

Here is the workflow this actually unlocks, and it rotates to inventory and restocking on purpose, because that is a task every physical business shares, whether you are running a hardware store, a boutique, a cafe watching its coffee bean stock, or a clinic tracking supplies.

Set up two agent roles instead of one. The first is your stock agent. It watches your inventory sheet or point-of-sale export, notices when something is running low, and drafts a purchase order: supplier, item, quantity, price. It does not send anything yet. It just drafts.

This is where mid-turn redirection changes the job. While the stock agent is still drafting, you can say, out loud or typed, "make that 20 units, not 50" or "switch to the backup supplier, the usual one is slow this month." The agent absorbs the correction mid-task and keeps going with the new instruction. You never had to stop it, re-explain the whole request, and hope it remembers the parts you already agreed on.

Once the draft is right, the stock agent hands the corrected order to a second role, a procurement agent, using the new A2A protocol. This second agent has one job: talk to the supplier, whether that is drafting the WhatsApp message, filling out a supplier portal, or sending the email, and it never touches your main inventory data. You have split "decide what we need" from "talk to the outside world," which is exactly the kind of separation that makes an automation safe to trust with real money.

Every time an order goes out, a signed webhook pings a Slack channel or a simple dashboard: what was ordered, from whom, and when. You get the paper trail without reading every message the agents exchanged.

Build It This Week: Four Steps

None of this needs a developer on staff. Here is the order to do it in.

1. Update Hermes. If you already run Hermes Agent through the desktop app or the Agent OS manager, go to Manage, then Update. If you have never set it up, Nous Research's Hermes is free and open source, and the desktop app is the easiest entry point.

2. Create two agent profiles, not one. Name the first something like "Stock," give it read access to your inventory sheet or POS export, and nothing else. Name the second "Procurement," give it access to your supplier contact list, WhatsApp Business, or email, and nothing else. Keeping their permissions separate is what makes the handoff safe.

3. Turn on the A2A plugin and register a webhook. Enable the bundled agent-to-agent plugin so the two profiles can discover each other. Then register a webhook endpoint, a Slack incoming webhook URL is the fastest option, and set it to fire on task completion and tool events.

4. Test the interrupt before you trust it with a real order. Ask the Stock agent to draft a purchase order for something low on your shelf. While it is mid-draft, correct a detail out loud if you have voice mode on, or by typing. Confirm it adjusts without losing the parts you had already agreed on, then let it hand off to Procurement and watch for the Slack notification. Only after that dry run should you let it touch a real supplier.

What This Costs You, And What It Buys You

Hermes Agent itself is free and open source. Your real cost is the underlying model's token usage, which for a small business running one or two agent profiles on routine tasks like this typically lands in the range of a few dollars to a few tens of dollars a month, depending on how chatty your workflow is and which model you connect it to.

What you get back is the thing that actually stops most small businesses from automating anything with real money attached: confidence that a mistake gets caught before it costs you, not after. A purchase order for the wrong quantity, sent to the wrong supplier, used to mean a phone call, an apology, and a week of waiting for the correction to ship. Now it means you said "wait, make that 20" while the draft was still open.

If restocking is not your bottleneck, the same three-part pattern, correct mid-task, hand off to a specialist, log the result, works anywhere you currently have one person doing two jobs: drafting a decision and then executing it. We have written before about building an AI agent that reorders stock before you run out, which covers the trigger logic for when to reorder in the first place. This piece is about what to do once the draft order is sitting in front of you and something about it is not quite right.