EREBOTS 4.0: A framework for conversational assistants with A2A agents and MCP tools

1. Motivation and significance

• A common platform holds discovery, transport, task lifecycle, streaming, identity, and conversation state fixed so researchers can vary and measure routing or delegation.The authors argue this makes results easier for others to replicate on the same substrate.Source: §1 · p. 3
• Earlier multi-agent platforms enabled composition and replication through specifications such as FIPA-ACL and XMPP, while current toolkits often leave user authentication and request-to-user linkage to applications.A2A and MCP now specify agent interaction and tool discovery or invocation, respectively.Source: §1 · p. 3
• EREBOTS 4.0 composes specialists over A2A and MCP across one process or separate hosts without code changes.A researcher lists agent endpoints in a configuration file; specialists can be written by subclassing a single class and naming their MCP servers.Source: §1 · p. 4
• Its application layer supplies OIDC login, persisted user context, execution traces, and one shared conversation service for frontends.The orchestrator propagates identity and user context to each local or remote specialist in A2A metadata.Source: §1 · pp. 3–4
• The paper reports measured conformance against the official A2A Technology Compatibility Kit, including untested requirements and two a2a-sdk deviations exposed by the measurement.The server layer overrides both deviations in the reported version.Source: §1 · p. 4

2.1. Software architecture

• A user turn streams routing decisions, answers, trace steps, and outcomes, while the orchestrator can send the stream directly in-process or as Server-Sent Events over a network.A contributing specialist’s output travels on a separate channel so the user receives a single answer.Source: §2.1 · p. 4, Figure 1 · p. 5
• The orchestrator discovers configured specialists from their agent cards, routes to one or more, and can call multiple specialists concurrently before synthesizing or concatenating their contributions.A failing call removes only its contribution; specialists receive the turn’s initial context and do not call one another.Source: §2.1 · pp. 5–6

2.2. Software functionalities

• The application supports identity propagation, persisted context and transcripts, task resumption, and local or independently hosted specialists, with shared connectivity and metadata conventions still required.Consent flags are stored and served to the application, which decides what to do with them.Source: §2.2 · pp. 7–8

2.3. Verification

• The automated suite passes 567 tests covering routing, A2A round-trips, task resumption, MCP behavior, authentication, concurrent persistence, and frontend events.The compatibility run uses fixture endpoints with canned payloads and no model in the loop.Source: §2.3 · p. 8
• 97 A2A Technology Compatibility Kit requirements passed with no failures across the reported runs; seven were skipped and 25 had no test in the suite.The 97 passed requirements are the union of two runs; the server layer overrides two a2a-sdk deviations to pass the relevant requirements.Source: §2.3 · pp. 8–9, Table 2 · p. 8

3. Illustrative examples

• Stubbed-reply tests pin routing, A2A calls, tool invocation, stored values, and task lifecycle, while specialist choice and asking for clarification remain model-dependent.The clarification reply updates the profile only after the user answers.Source: §3 · p. 9
• A parked task resumes after a server and store restart only for the same identity and before expiry.The figures show a run with persistent tasks enabled; the automated suite exercises the same mechanism with stubbed replies.Source: §3 · p. 10
• In the clarification run, PersonalAgent asks for a missing value in Italian, parks the task as input-required, then completes after the Spanish answer updates the profile.The client view carries traces and timings on every turn; the database view shows the task state and profile change.Source: §3 · p. 9, Figure 2 · p. 10, Figure 3 · p. 11
• Fan-out routes independent profile and conversion operations to both agents, returns 19.312128 kilometres for 12 miles, and preserves the profile result if conversion is unreachable.Reproducing persisted outcomes requires Keycloak and persistent tasks; measuring model-dependent behavior requires a pinned model, fixed prompts, and a chosen repetition count.Source: §3 · p. 11

4. Impact

• EREBOTS 4.0 makes routing, delegation, and specialist granularity configurable study variables by fixing discovery, streaming, task lifecycle, and authentication.The mechanics are verified with stubbed model replies, so configurations can differ in the component under study and the model.Source: §4 · p. 12
• Official A2A conformance measurement reports untested requirements, and shared fixtures let groups check servers and compose agents without shared codebases.Two groups that meet the specification can compose each other’s agents.Source: §4 · p. 12
• Self-hosting lets specialists run where their data and tools already reside, while operators decide data location and jurisdiction.The framework needs an OpenAI-compatible model endpoint, which can be served by a local inference server.Source: §4 · p. 12
• Code-enforced identity and context propagation, deterministic MCP results, specialist failure isolation, resumable owner-bound clarification tasks, linked traces, and consistent shared transcripts do not depend on model behavior.Persisted clarification tasks resume across restarts when enabled.Source: §4 · p. 12
• Current boundaries include one queued clarification, unenforced consent flags, uncoordinated replicas, no specialist negotiation, and unmeasured model-dependent routing and ask-user success.Persisted tasks survive restarts of a single instance, but replicas are not coordinated.Source: §4 · pp. 12–13

• 97 requirements of the official A2A conformance suite passed with no failures.Source: §5 · p. 13
• The user agent serves an agent card, while the orchestrator routes requests to matching MCP skills and composes specialist responses.Source: §5 · p. 13
• The framework stores consent flags for applications but does not refuse operations when consent is absent.Source: §5 · p. 13
• A consent decision for a remotely hosted specialist cannot travel with the request because A2A has no field for it.Source: §5 · p. 13
• Specialists can call one another as A2A servers; loop detection and hop limits in message metadata could bound those exchanges.Source: §5 · p. 13
• The paper leaves open whether agent negotiation outperforms orchestrator routing, noting that LLM bids cost model calls and suitability outputs may be wrong.Source: §5 · p. 13

• EREBOTS 4.0 is distributed as source code and a Python wheel.Source: Current executable software version · p. 17
• An optional Docker Compose stack is available for the full deployment.Source: Current executable software version · p. 17
• No standalone executable is distributed, so the optional software metadata table does not apply.Source: Current executable software version · p. 17