Repository navigation
Replies: 9 comments
|
Created ticket #2134 for the issue mentioned under Known Limitation, and being able to replace an agent or undeploy an agent so that stale cards aren't visible to the planner. |
|
@AlBaker - Thank you for introducing Agent Spaces to Embabel. May I suggest starting with the Discussion Forum: https://github.com/embabel/embabel-agent/discussions Meanwhile, I will review the Agent Spaces architecture. Thank you |
added as a sub-issue, thanks |
|
@AlBaker, could you please elaborate on your vision for the INSIDE integration framework? Per my understanding, should it not rather be the OUTSIDE framework? thanks |
|
@AlBaker - will be moving to discussion, this is advanced notice, thank you |
|
I'm okay with discussions, though haven't used that feature before. I presume you'll create a matching entry in github discussions and cross-link here? Additionally, we're working on an architecture guide set of documents to go with AgentSpaces. I'm curious how you view Spring AI and Embabel. Supporting both, I kind of want Embabel to be higher level, help me define the agent, pull data/store results in AgentSpaces, and generally abstract away a lot of the underpinnings of the chat. This is actually where we started, and you'll probably see the original concept was "let's have Embabel use Jini to save its results, so other agents can check/use those results." When I'm working with Spring AI, I kind of zoom in a bit, and I'm putting prompts/results and such in the space so that other agents can re-use my chats, or any piece there of. We may find some token-reduction value here, but the application developer has rolled up their sleeves compared to working with Embabel. Probably a sign that it'd be good for us to hash it out in the discussion forum. I'll post a note accordingly when we push the architecture guide to the website. |
|
@igordayen FYI architecture guide is up: https://www.badmonkey.ai/agentspaces/developer-guide/architecture-guide.html |
|
I'm curious how you view Spring AI and Embabel. @AlBaker, Embabel uses Spring AI as the default LLM connector, but it is built with Converted to Discussion. |
|
@igordayen I would wager we view them the same, in one way, Embabel layers over Spring AI and you get all the joy of Spring apps available to Embabel agents, but you can skip that and just change out the LLM. AgentSpaces is nearly the old Jini playbook, with a bit more aligned for agents. So it's a valid tuple space and suite of services that you could use at either of those layers, the thing that changes is what you put in the space (agent/fleet actions/state, business objects, LLM chat sequences, etc). The embabel-agentspaces bridge, for example, extends the GOAP planner to see all of the agents across the peer network, so you can include them in its planning. I saw today that you guys are headed down the graph road. I spent 13 years as VP of Enterprise Solutions at Stardog, leading graphs builds, probably having built more of them than anyone. 15 years ago, I wrote the stardog-spring bindings, a simple template API using Spring conventions, auto-wiring and the rest of it, and got super popular. |
Uh oh!
There was an error while loading. Please reload this page.
Summary
Bad Monkey proposes an AgentSpaces subsection under Integrations, beside A2A, describing the embabel-agentspaces bridge. The bridge publishes each
@Agent's metadata as a signedAgentCardinto a peer group, and turns the group's cards into a typed@Agentwhose@Actionmethods invoke those remote capabilities. So Embabel's GOAP planner reaches goals whose actions run on other machines without changing how an Embabel application is written. We tested it against Embabel 1.5.2 on Spring Boot 4.1 in a real-platform integration test, and Party Bus runs Tripper's planner over a 34-agent fleet with it.You can see the full sample app in the agentspaces-partybus repo.
Why this fits Embabel
Embabel's documentation says agent systems will support federation, with planning that incorporates remote actions and goals, both with other Embabel systems and with third-party frameworks. An agent process and its blackboard live in one
AgentPlatformtoday. The A2A client (#435) gave Embabel a way to discover and invoke remote agents one call at a time, and #1479 asks for a tool set that changes while an agent runs. AgentSpaces offers the replicated TupleSpace as a way to extend that blackboard to other systems, enabling Embabel agents to share that blackboard with LangChain, Micronaut, or even just other instances of Embabel in the peer-to-peer fabric.The bridge addresses the same integration from the planner's side. Every capability advertised in the peer group arrives as an
@Actionmethod the planner can chain, condition on by type, and replan around. When a peer joins or its card lapses, the generated agent is redeployed with the new action set, so the planner's world grows and shrinks with the fleet.What AgentSpaces is
AgentSpaces is a peer-to-peer coordination fabric for agents, services, and the people who work with them. Peers join a group, each with a cryptographic identity (ie PKI). Participants advertise what they can do through signed, leased documents: agents publish AgentCards, and databases, sensors, and services publish AssetCards. All of them share a replicated tuple space with
write,read,take, andnotify. Basically, providing the TupleSpace from good ole Jini. However, Jini required a centralized broker, so we've also resurrected parts of Juxtapose (JXTA), for discovery, routing within the peer group, including supporting relay and rendezvous. All done with modern TLS specs, all open source and Apache 2.The project is
github.com/badmonkeyai/AgentSpaces, Apache 2.0. The embabel bridge lives in the main repo, with separate repos for Spring AI, langchain4j, etc.How the bridge works
Three pieces, all in
embabel-agentspacessee here:EmbabelBinderreads each@Agentbean's metadata reflectively and publishes it as AgentCards through the AgentSpaces binder, so the cards are leased and refreshed like every other card. It skips beans the space binder already handles.EmbabelRemoteActionswatches the group's advertisement cache and generates (with Byte Buddy) a typed@Agentclass with one@Actionmethod per foreign card and consumed/produced type pair. The methods are real because the planner conditions on method signatures. Invoking one writes the request entry into the space and awaits the correlated result. This is choreography over the space: whichever peer takes the request answers it, and a peer that dies mid-way hands the request to the next.EmbabelRemoteActionsDeployerdeploys that agent onto theAgentPlatformthrough Embabel'sAgentMetadataReader, and redeploys it when the set of actions changes. A refresh that changes only a card's issue time triggers no redeploy. The property isagentspaces.embabel.auto-deploy, on by default.The module reads Embabel reflectively, so it builds with no Embabel artifact and activates only when Embabel is on the application classpath. An Embabel application adds the AgentSpaces starter and this module, declares a group in
agentspaces.groups, and keeps its own agents as they are.Integration Examples
EmbabelFleetPlanningTestin the AgentSpaces repository runs against the real Embabel 1.5.2 artifact:A worker on a plain AgentSpaces peer, with no Embabel on its classpath, advertises "Brief in, Summary out":
A Spring Boot 4.1 application with Embabel and the bridge joins the same group. The worker's card reaches its discovery cache, and the deployer puts the generated fleet agent onto the real
AgentPlatform.Embabel's planner plans
BrieftoSummaryand runs the fleet action:The remote worker performs it, and the test asserts the result.
The goal type is one the planner provably cannot reach with its local actions. It becomes reachable, and is reached, the moment the fleet's cards are in the action set.
Party Bus: Tripper with a Fleet of Agents
If you have a large family like I do, there's lots of competing interests, attempts to gain quorum, restrictions on who can do what, etc. So I thought this would be a great vacation simulator that drives the fleet through a week in Lisbon. Everything that can go wrong, goes wrong on purpose: a storm re-auctions a day, an illness collapses the party's energy and the rest planner's bid takes the day, and so on. We thought this would be a great twist on Tripper, demonstrate why you bring AgentSpaces into the mix, and showcase both libraries working together.
Party Bus takes the Embabel travel planner, Tripper, and keeps its domain model and its planner, while handing the work to a fleet of 15+ agents on many peers who coordinate through shared spaces. Every coordination type appears once, exercising different parts of AgentSpaces and Embabel together.
@SpaceTakeon a research bag under lease race, kill-tolerant@BidFunctionand@SpaceTakeon theslotsspace under auction, ie using bid and take on the space as the coordination@SpaceTakesweeps,@SpaceNotifyscoring, signed quorum voteThe project is
agentspaces-partybus, and although there are many agents, many peers, it runs on your laptop under a central Maven build.Proposed documentation content
A subsection under Integrations, parallel in shape to the A2A section:
embabel-agentspacesdependencies, a minimalagentspaces.groupsblock, and theauto-deployproperty.We will write the subsection and open the pull request if we get agreement on this ticket.
Known limitation
The
AgentPlatforminterface the module reaches exposes no undeploy, so an action whose card has vanished stays in the previous generation of the generated agent until the platform restarts. The deployer redeploys with the new set, but the old generation remains registered. We'll create a separate issue for an undeploy or replace operation.What it does not do
The Asks
embabel-agent-docsmodule, so we can author content accordingly.Bad Monkey, Inc. authored the AgentSpaces core library and corresponding Embabel bridge. We're happy with whatever level of engagement you'd like, whether simply growing the open source ecosystem or discussing potential commercial relationships.
All reactions