Hermes Agent v0.11 Just Dropped — This Changes Everything

News Research just dropped Hermes agent

version 0.11. And this one is not a

small update. Since version 0.9, this

project has seen 1,556

commits, 761 merged pull requests, and

contributions from 29 community

developers. The release notes call it

the interface release, and that framing

is accurate. This is not a patch. This

is a structural overhaul across three

layers of the entire stack. the

interface you interact with every day,

the infrastructure running underneath

it, and the surface area for what you

can actually build on top of it. Now, if

you have been following this series, you

already know that version 0.10 shipped

only the news tool gateway and deferred

almost everything else. What that means

is version 0.11 carries roughly 2 weeks

of accumulated work across the full

codebase, all landing at once. The TUI

has been completely rewritten. The

transport layer has been pulled out into

a pluggable architecture. AWS Bedrock

now has native support. Five new

inference providers are in. GPT 5.5 is

accessible through your existing codeex

OOTH. The plug-in surface has expanded

significantly and there is a new command

called / steer that changes how you

interact with a running agent mid task.

We're going to cover all of it and we're

going to spend serious time on the

agentic workflows that these new

capabilities actually unlock. Before we

get into it, if you're finding value in

this series and you want to stay current

on everything happening in the open-

source AI agent space, hit the like

button and subscribe. It genuinely helps

this channel grow and it makes sure you

do not miss the next breakdown when

these projects keep moving this fast.

Share this one with anyone in your

network who is building with AI agents

or evaluating open source tooling

because this release is worth knowing

about. The most visible change in

version 0.11 is the interactive CLI. The

entire thing has been rewritten from

scratch using React and inc land across

roughly 310 commits. You launch it with

Hermes--TuI

where you can set the environment

variable Hermes_TUI

to one if you want on by default. What

you get when you open it is a

fundamentally different experience from

the classic terminal interface. The

composer is now sticky. It stays fixed

at the bottom of the screen while your

history scrolls independently above it,

which sounds like a small thing until

you're deep in a long session and you

realize you're no longer fighting the

terminal to keep your input visible.

Streaming output renders live as the

agent works. And OSE52 clipboard support

means copy operations work correctly

even when you are connected over SSH.

That one has been a quiet painoint for

anyone running Hermes agent on a remote

VPS and it is now handled properly out

of the box. The status bar has been

extended with two additions that are

actually useful during active runs. Your

current git branch now shows in the

status bar alongside your working

directory. So you always know which

branch the agent is operating on without

having to run a separate command. And

there is now a per turn elapse stopwatch

that counts up while the agent is

working and shows you the total time

when the turn completes. When you are

running multi-step tasks and trying to

understand where time is going, having

that number visible in the interface is

far more useful than guessing. There's a

new command in this release called /

steer, and it deserves more than a

passing mention. The problem it solves

is one that anyone who runs long agent

tasks will recognize immediately. You

kick off a task, the agent starts

executing, and somewhere a few tool

calls, and you can see it is going a

direction you did not intend. Before

this, your options were to interrupt the

run and restart with a corrected prompt

or let finish and clean up afterward. /

steer gives you a third option. You type

/ steer followed by your correction and

the agent receives that note after its

next tool call completes. It does not

interrupt the current turn. It does not

break the prompt cache and the task

continues from where it was with your

correction folded in for long running

research tasks, code generation

workflows or anything where restarting

has a real cost. This is a significant

quality of life improvement. The release

also ships a sub aent spawn

observability overlay in the TUI. When a

parent agent delegates work to child

agents, you can now watch those spawns

happen in real time inside the

interface. If you're building or

debugging multi- aent workflows, having

that visibility without adding your own

logging is exactly the kind of thing

that shortens feedback loops. A light

theme preset and a/clear confirmation

dialogue round out the UX changes. And

the whole thing is backed by a Python

JSON RPC backend called TUI gateway that

keeps the interface layer cleanly

decoupled from the agent runtime. One of

the most important changes in version

0.11 is one that most people will scroll

past in the release notes and that is

the transport layer ref actor.

Previously, the logic responsible for

format conversion and HTTP transport was

embedded directly inside run agent. py.

It worked but meant that adding a new

provider or changing how requests were

shaped required touching core agent

code. In this release, all of that has

been extracted into a dedicated

agent/transports

directory with a clean abstractbased

class and four concrete implementations

sitting on top of it. Those four

transports are anthropic transport for

the anthropic messages API, chat

completions transport as the default

path for any open AAI compatible

provider, responses API transport for

the open AI responses API and codecs

build cores wiring and bedrock transport

for AWS bedrock via the converse API.

Each one owns its own format conversion

and API shape independently. The

practical result is that the agent

runtime no longer needs to know anything

about how a specific provider

communicates. That concern now lives

entirely inside the transport layer,

which makes the whole system

significantly easier to extend going

forward. The headline edition built on

top of this new architecture is native

AWS bedrock support. And it is worth

spending a moment on why this matters

beyond just having another provider

option. When you route Hermes agent

through Bedrock transport, your traffic

flows through your own AWS environment

using the Bedrock commerce API. It does

not leave your infrastructure to hit an

external endpoint. For teams operating

under data residency requirements,

enterprise security policies or

compliance frameworks that restrict

outbound API traffic. This changes the

equation entirely. You get the full

capability of Hermes agent including its

tool system, memory, multi- aent

delegation and messaging platform

integrations while keeping all model

traffic inside the boundary your

security team has already approved.

Getting set up with bedrock transport

follows the same configuration pattern

as any other provider in Hermes agent.

You define your provider in the config

pointed at your bedrock converse API

endpoint. Supply your AWS credentials

through the standard credential chain

and select whichever Bedrock hosted

model you want to run against. The

transport handles everything from there.

And because the transport layer is now

properly abstracted, switching between

bedrock and any other provider in your

setup is a config change, not a code

change. That kind of operational