Repeated discovery
Where is the right page? Which field comes next? Agents spend effort rediscovering how an application works.
SMART365GUIDE ASDK · AGENTIC SDK
No more
guessing.A clear path
to done.
Give AI agents a clear path to completing work inside your software.
Application knowledge. Approved workflows. Permission checks. Verification of the final result. Your software, with your team in control.
Built for compatible agents
Codex Claude ChatGPT Local models
Authorized browser tools or a configured adapter required.ONE APPLICATION. A CLEAR PATH.
THE AGENT’S GOAL
REQUESTCreate a purchase order.
ILLUSTRATED STEP 1 / 6
Relevant application context selected
Interactive architecture illustration. No live agent or business action.
APPLICATION KNOWLEDGEAPPROVED WORKFLOWPERMISSIONVERIFIED RESULT
THE PROBLEM
Software was built for people.
Agents are still finding their way.
An agent often has to discover where to go, which fields to fill, what actions are available and whether its work was saved. Complex workflows make every wrong turn more expensive.
Retries and extra requests
Wrong turns can consume model tokens, generate additional application requests and delay completion.
Unclear control and outcomes
Software owners need to control agent access, understand what happened and confirm that changes were saved.
WHAT WE’RE BUILDING
Your software knows the way.
Give the agent a guide.
Smart365Guide is a guidance and execution layer between your application and a compatible AI agent. Supply relevant context, clearly defined actions and an approved path through the workflow.
Your application
Routes · fields · backend tools
Your permissions stay authoritativeSmart365Guide ASDK
Context → approved path → checked result
Your team decides what is availableCompatible AI agent
Codex · Claude · ChatGPT
Local tool-calling models tooHOW IT WORKS
From a request
to a checked result.
Connect what your application knows to what an agent is allowed to do. Each stage has a purpose. Your backend remains the source of truth.
01 / CONTEXT
The right context. At the right moment.
Your team embeds the SDK and supplies relevant routes, fields and reviewed workflows. Give the agent the application knowledge needed for its goal.
PURCHASE ORDER · REFERENCE EXAMPLEGUIDED PATH
Create a purchase order.
Application contextRequired inputs selected
Action boundaryPublished capability
Current stageApplication knowledge
Relevant application context selected
Illustration only. Actual actions need a configured provider or authorized browser tools. Guidance alone does not execute or verify work.
A GUIDE FOR PEOPLE, TOO
The same knowledge. A helpful conversation.
An embedded conversational guide helps people understand the application, find the right page and follow reviewed workflows. Questions, navigation and assistance stay in one chat.
RECORDED PRODUCT DEMO
The real bot. The real application.
A project that actually gets created.
Watch this Codex session use published ASDK guidance and browser controls to register an SDK application, check the saved result and inspect billing in Smart365Guide.
CODEX SESSION + SMART365GUIDE ASDKRecorded in a local test workspace
Application saved and read backSDK credentials generated; values hiddenBilling inspected · no payment
Actual SDK, authenticated MCP guidance and application APIs, operated by the Codex assistant in this session. Synthetic account and invoice. This is not an external Codex CLI run or a test of ChatGPT or Claude. Pauses are trimmed and frames may be held for the AI-generated narration. No speed or token measurement is claimed.
Recording evidenceUSE YOUR OWN AGENT
Connect the client.
Keep your account in control.
Sign in, publish the workflow for the integrated application, and create a delegated agent key in ASDK setup. The agent also needs authorized browser controls or an explicitly granted backend action.
Open ASDK setupKeep agent keys in the client’s private runtime. Do not paste passwords or agent keys into prompts, public pages or the browser embed.
CLI CONFIGURATION · BROWSER TOOLS ALSO REQUIRED
Add the ASDK MCP connection
Set S365_NETWORK_KEY privately in the environment that launches Codex. Then add the server and restart the client. The recording above uses this assistant’s browser tools, not a separate CLI invocation.
codex mcp add smart365guide-asdk --url "https://www.smart365guide.com/api/v1/asdk/mcp" --bearer-token-env-var S365_NETWORK_KEY
EXAMPLE REQUEST · AFTER CONNECTION AND WORKFLOW SETUP
Use the published Smart365Guide ASDK workflow to create an SDK application named My Web App for app.example.com in my signed-in workspace. Confirm the saved name, allowed domain and SDK credential presence without revealing keys. Then read my current plan and invoice status. Stop before any checkout, upgrade or payment. If the published path, authorization or browser tools are missing, explain what is needed instead of claiming success.
What this demonstration proves
A published workflow was read over real authenticated MCP. The embedded SDK displayed the reviewed guide. Browser controls filled and submitted the actual registration form once, which returned HTTP 201. A separate authenticated read confirmed the saved project, allowed domain and credential presence. Billing was read from a synthetic workspace invoice. No checkout request or real charge occurred, and the temporary agent key was revoked after the recording.
EARLY EVIDENCE
Less discovery.
Measured, not assumed.
Our controlled Codex pilot compared browser-only work with a guided integration in a synthetic web application, including SDK startup in the task clock.
MEAN TOTAL TOKENS
48.4%lower token usage
Normalized mean token indexMEDIAN CONFIRMATION TIME
15.9%shorter completion time
Normalized median independently confirmed timeCONTROLLED CODEX PILOT
Pilot methods & recorded evidenceWorkload-specific results, not a guarantee for every agent or application. The assisted setup combines the SDK guide, registered backend capabilities and an adapter; it does not isolate the SDK alone. Lower tokens do not imply the same percentage of billed savings.
WHAT WE’LL MEASURE NEXT
RetriesApplication requestsInfrastructure useCost per successful task
Improved task success and greater load capacity are benefits to validate. Agent operators may pay the model bill; software owners pay hosting and backend costs. We aim to make both sides measurable.
WHY NOW
Agents are taking action.
Software needs a way to respond.
As agents move from answering questions to working inside business software, teams need a practical way to expose capabilities, enforce permissions and verify outcomes.
THE WORKFLOW CHALLENGE
A goal needs an actionable path.
A request to update a record or create a purchase order can cross pages, forms and approval steps. Publish the relevant inputs and reviewed path so a compatible agent can follow your application’s workflow.
See the guidance contractTHE SHIFT TOWARD ACTION
Taking action needs clear boundaries.
Publishing application knowledge does not grant access. Keep account roles, delegated scopes, approvals and saved-state checks attached to the work. Your software owner decides what an agent may do.
Explore approved capabilitiesWHO WE’RE STARTING WITH
One useful workflow.
A measurable starting point.
Our proposed initial focus is B2B SaaS companies with repetitive customer-service and operations workflows. Start in staging, compare the existing approach and expand based on results.
CRM updates
Update the right record. Preserve the rest.
Check the requested changes against the saved customer record and the user’s permissions.
Service requests
Route, update and confirm the request.
Guide required inputs, respect approval points and check the final status.
Purchase orders
Create it. Read it back. Confirm it.
Collect the required inputs, use the approved action and verify the saved purchase order.
01Choose a bounded task
02Integrate in staging
03Compare real outcomes
04Expand with evidence
THE SOFTWARE OWNER STAYS IN CONTROL
A clear path
to done.
We’re building toward software that can explain its capabilities to agents, authorize their actions and verify their work.
Web prototype · compatible agents · owner-approved integrations