AI DevCraft · Live webinar
30 September 2026 · 11:00 am MT

Agents All the Way Down:The Agentic Stack

Software development · Backend · Frontend · Browser

Rainer Hahnekamp & Murat SariSoverius AI
1
Soverius AI — We build production-ready AI systems
Soverius AI

Agenda

Four chapters and their relationships Agentic coding encompasses backend agents, agentic UI and browser-local agents. Agentic UI connects to backend agents. Backend agents and browser-local agents both depend on models, shown outside the coding frame as context. 01 Agentic coding 02 Backend agents Mastra · LangGraph · Google ADK · CrewAI 03 Agentic UI A2UI · AG-UI · CopilotKit 04 Browser-local agents WebLLM · Prompt API · WebMCP Models Closed · Open weights
3
Soverius AI

01 · Agentic coding

Current chapter: Agentic coding Agentic coding encompasses backend agents, agentic UI and browser-local agents. Agentic UI connects to backend agents. Backend agents and browser-local agents both depend on models, shown outside the coding frame as context. 01 Agentic coding 02 Backend agents Mastra · LangGraph · Google ADK · CrewAI 03 Agentic UI A2UI · AG-UI · CopilotKit 04 Browser-local agents WebLLM · Prompt API · WebMCP Models Closed · Open weights
4
Soverius AI

Agents in the loop

A coding agent becomes one step in a framework-orchestrated workflow alongside triage and review agents. Humans remain involved; loop development requires substantial financial resources or local models. 5
Soverius AI

Implementation Loop

The implementation loop in four phasesIssue creation and agent triage, optional developer confirmation, sandboxed implementation, coordinated agent review with continue, fix or human escalation, then a pull request and review report for a human merge decision. Fixes and requested changes return to implementation. fixchanges requested 01 Issue & triage Create issue user intent + context Triage agent is the spec complete? Developer confirm specification OPTIONAL Missing detail? Ask the author. 01Issue & triage 02 Sandboxed implementation SANDBOX Worker assigned specification Implement isolated worktree Check tests + evidence Explicit tool, file and network permissions 02Sandboxedimplementation 03 Agent review & decision Review agentsRUN IN PARALLELPerformanceArchitectureVisualsTest code qualityBugsMain reviewagentssummary + reportDecisionContinue → PRFix → phase 2Ask a human…Jev03Agent review& decision 04 Pull request & human review Pull request change + agent report Human review Merge Do not merge 04Pull request& human review Improve the loopOptional approvalOptional escalationRequired review🧑‍💻Human🤖AI agent
6
Soverius AI

02 · Backend agents

Current chapter: Backend agents Agentic coding encompasses backend agents, agentic UI and browser-local agents. Agentic UI connects to backend agents. Backend agents and browser-local agents both depend on models, shown outside the coding frame as context. 01 Agentic coding 02 Backend agents Mastra · LangGraph · Google ADK · CrewAI 03 Agentic UI A2UI · AG-UI · CopilotKit 04 Browser-local agents WebLLM · Prompt API · WebMCP Models Closed · Open weights
7
Soverius AI

An agent use case: intent to SQL

Human intent

A question in natural language

Agent

Generate SQL from the schema

SQL query

Check, execute, show results

    8
    Soverius AI

    Beyond predefined filters

    “Show the highest air temperature for each shift manager.”

    • The user describes the result they want
    • The agent generates a query for that intent
    • Application code checks and executes the query
    9
    Soverius AI

    The SQL workflow

    Generate SQL

    Model proposal

    Deterministic check

    Statement policy

    Agentic check

    Jev

    Run query

    Read-only database

      10
      Soverius AI

      Review SQL with Jev

      Safety

      Is the query read-only and allowed?

      Intent

      Does it answer the question?

      Columns

      All and only the expected columns?

      jev-sql-reviewer.ts · selected source
      const questions = {
  safety: createSafetyCheck(),
  intent: createIntentCheck(),
  columns: createColumnCheck(),
};
      11
      Soverius AI

      03 · Agentic UI

      Current chapter: Agentic UI Agentic coding encompasses backend agents, agentic UI and browser-local agents. Agentic UI connects to backend agents. Backend agents and browser-local agents both depend on models, shown outside the coding frame as context. 01 Agentic coding 02 Backend agents Mastra · LangGraph · Google ADK · CrewAI 03 Agentic UI A2UI · AG-UI · CopilotKit 04 Browser-local agents WebLLM · Prompt API · WebMCP Models Closed · Open weights
      12
      Soverius AI

      A spinner is not enough

      Incident assistantHelps investigate facility incidents

      Check the air temperature in Packaging hall.

      Thinking… 0:00

      Done! Alarm raised.

      Raised an alarm for Packaging hall and notified the operator.

      Tool callsquery_historianraise_alarm
      1. What did it just do?
      2. Who said yes?
      3. How would you have stopped it?
      SEEASKDECIDEDELEGATE

      only a spinner

      never asked what to do

      no approval before raising the alarm

      an action with no oversight

      The human decision is missing.

      13
      Soverius AI

      A run is more than a model response

      Thinking… What the user sees
      USER TASK
      1. request you ask
      2. model call thinking one model response
      3. tool call query_historian
      4. ask the human 🧑‍💻 you choose
      5. model call propose alarm
      6. approval interrupt
      7. execute tool record outcome
      8. answer summary

      The UI needs progress, human input and results—not just the model response.

      14
      Soverius AI

      One contract for the interaction

      PredictThree agent backends, two UIs, and you add approval. How many pieces of work?

      AG-UI defines the interaction protocol.

      Implemented withCopilotKit
      15
      Soverius AI

      Three typical Agentic UI use cases

      01 Frontend tool call

      Change the view through chat

      “Go back to the snapshot view.”
      SnapshotReading log

      02 Activity progress

      See the workflow advance

      Reviewing SQL with Jev50%
      0255075100%

      03 Human in the loop

      Approve before raising an alarm

      🧑‍💻 Raise an alarm?

      Packaging hall · Air temperature

      RejectApprove
      16
      Soverius AI

      04 · Browser-local agents

      Current chapter: Browser-local agents Agentic coding encompasses backend agents, agentic UI and browser-local agents. Agentic UI connects to backend agents. Backend agents and browser-local agents both depend on models, shown outside the coding frame as context. 01 Agentic coding 02 Backend agents Mastra · LangGraph · Google ADK · CrewAI 03 Agentic UI A2UI · AG-UI · CopilotKit 04 Browser-local agents WebLLM · Prompt API · WebMCP Models Closed · Open weights
      17
      Soverius AI

      AI in the browser: technology options

      TechnologyModel managementExecution
      Prompt APIBrowser-provided modelBrowser downloads and manages itChrome: Gemini NanoBrowser-managedCPU or GPU in Chrome
      WebLLMLLM runtimeYour app selects a compatible modelWebGPUGPU
      ONNX Runtime WebUsed by Transformers.jsYour app selects a compatible modelWebAssembly or WebGPUCPU or GPU

      Our demo: Transformers.js → ONNX Runtime Web · E5 + Gemma

      18
      Soverius AI

      Agents inside the browser

      Download

      Compatible model

      Load

      Transformers.js

      Run

      On-device inference · WebGPU

        19
        Soverius AI

        Why RAG?

        01

        Internal documents

        Use your own knowledge.
        Link answers to their sources.

        Manuals · policies · procedures
        02

        Context when needed

        Retrieve relevant passages
        when the user asks.

        Selected context, not the whole library
        03

        Similar meanings

        Find relevant text even when
        the wording differs.

        “The cold room warms up.”↕“Room temperature too high.”
        20
        Soverius AI

        RAG in the browser

        Both models: Transformers.js → ONNX Runtime WebE5: WASM / CPU · Gemma: WebGPU / GPU

        🧑‍💻 USER QUESTION “The cold room warms up.” MANUAL PASSAGE “Room temperature too high.” UNRELATED EXAMPLE “The printer is out of paper.” E5 Embedding model VECTORS [0.80, 0.60, 0.00] [0.76, 0.64, 0.12] [−0.50, 0.20, 0.84] Illustrative values · real E5 vectors: 768 dimensions COSINE SIMILARITY 0.99 Closer match −0.28 Different meaning QUESTION + SELECTED PASSAGE TEXT “The cold room warms up.” Manual: check door seals; keep doors closed. [p. 6, 2.6] Gemma Language model ANSWER TO THE USER Check the door seals and keep doors closed. [p. 6, entry 2.6]
        21
        Soverius AI

        Agents everywhere

        • Development: implement and check features
        • Backend and UI: investigate with the user
        • Browser: local inference and selected actions
        22
        Soverius AI

        Thank you!

        We would love your feedback.

        soverius.ai/feedback/7654629a

        Demo code · Units 2, 3 & 4

        github.com/Soverius-AI/
        talk_ai-dev-craft_agents_agentic-stack
        QR code to https://soverius.ai/feedback/7654629a
        Scan to leave feedback
        23