Founder-led product · nine-month build cycle
Agent Mag
A product family for AI-agent builders and operations teams: an 86-page public platform, a 492-skill open registry and CLI, Agent Mag AI, and a governed automation runtime whose first deep domain is finance operations.
The demo is access-controlled. Architecture and scale describe the reviewed platform; credentials and source access are shared privately when appropriate.
86
product pages in the reviewed platform
62
API routes across product domains
16
CI, release, sync, and review workflows
492
public installable skill bundles
The system
From operational problem to reliable workflow.
Problem
High-stakes operational work crosses finance tools, inboxes, documents, spreadsheets, and approval chains. A useful platform has to do more than generate an answer: it must separate probabilistic reasoning from controlled execution, preserve state, pause for review, recover from failure, and leave an audit trail.
Approach
Harshitha designed Agent Mag as connected product layers rather than a single chatbot: editorial and discovery surfaces for builders, a declarative skills registry and npm CLI, an OpenClaw-backed agent interface, and a governed runtime where agents investigate while typed workflows, approvals, and durable state control consequential actions.
Outcome
The result is a credible product and distribution system around agent infrastructure: 86 product pages, 62 API routes, 16 delivery workflows, 492 public skill bundles, and a reusable operations foundation that can extend from finance into broader back-office work.
Architecture
System architecture
Agent Mag is Harshitha's founder-led product family. The public platform helps builders discover technical content, tools, models, jobs, events, resources, and installable agent skills. Agent Mag AI provides a tool-using interaction layer. A separate governed automation runtime applies the same agent-building discipline to operational workflows where the surrounding system must own permissions, execution order, approvals, integrations, and evidence. Finance is its first substantial domain, but the primitives extend to contract, customer, inventory, document, and internal-service workflows.
Builder platform
Unifies editorial, tools, models, skills, jobs, resources, events, accounts, partner surfaces, and documentation into a coherent product for AI-agent builders.
Next.js 16 · React 19 · 86 pages · 62 API routes
Agent Mag AI
Streams model responses, executes bounded function calls, renders structured tool results, and routes through an OpenClaw workspace with public-user identity and security constraints.
Azure OpenAI · OpenClaw · function calling · streamed UI blocks
Open ecosystem
Distributes declarative skill bundles and builder tools through a public registry, product catalog, APIs, GitHub repositories, and an npm CLI.
492 skills · Agent Mag CLI · GitHub Actions · npm
Operator Studio
Composes workflows on a visual canvas and gives operators access to agents, runs, files, knowledge, tables, templates, schedules, dashboards, and deployment state.
Next.js · React · XYFlow · Zustand · TanStack Query
API control plane
Validates requests, applies authentication and tenancy context, manages configuration, receives webhooks, and starts or observes execution without keeping long work inside the request lifecycle.
Hono · TypeScript · Zod · Better Auth · Socket.IO
Durable execution
Runs long-lived tasks, agent steps, extraction, browser work, transformations, approval waits, and provider actions with retryable worker semantics.
Trigger.dev · TypeScript workers
Operational state
Stores workflows, deployments, runs, approvals, audit events, files, knowledge, company-scoped records, and integration metadata as reconstructable state.
PostgreSQL · Drizzle ORM · 31 migrations
Module and integration layer
Exposes typed core blocks and provider adapters for email, documents, accounting, CRM, payments, inventory, signatures, storage, and collaboration systems.
16 core blocks · 17 integration modules · Nango
Reasoning layer
Uses models for investigation, extraction, classification, drafting, and agent decisions while returning bounded outputs to the workflow runtime.
OpenAI · Anthropic · structured schemas
Execution model
End-to-end execution
- 01
Compose and validate
An operator assembles typed blocks, connects integrations, supplies company context, and validates required configuration before deployment.
- 02
Start a durable run
The control plane creates persistent run state and hands long-running work to the worker runtime instead of tying it to one browser or HTTP connection.
- 03
Investigate with bounded agents
Agent steps can search knowledge, inspect documents, extract structured fields, or prepare a proposed action without receiving authority to execute every downstream effect.
- 04
Pause for human control
Approval blocks suspend consequential work and expose the proposed inputs and context for an operator decision.
- 05
Execute deterministic actions
Approved provider operations run through typed integration modules such as accounting, CRM, document, email, payment, or inventory adapters.
- 06
Persist evidence
Outputs, errors, timestamps, approval decisions, and audit events remain connected to the run and company so an operator can reconstruct what happened.
Implementation
What Harshitha implemented
- Implemented real Azure OpenAI streaming and expanded the chat surface into a multi-step tool loop with function calls and visual result blocks.
- Connected Agent Mag AI to an OpenClaw gateway, added explicit model routing, and refactored the surface around a dedicated OpenClaw execution path.
- Hardened the public agent with identity locking, untrusted-user rules, internal-detail boundaries, tier-aware behavior, persona files, and workspace context.
- Shaped an 86-page product spanning editorial, tools, models, skills, jobs, events, resources, accounts, partners, documentation, and AI-native discovery.
- Built the open ecosystem around 492 declarative skill bundles, CLI installation, public tools, documentation, and automated synchronization and release paths.
- Mapped operational procedures into explicit triggers, model-assisted investigation stages, deterministic actions, waits, conditions, parallel branches, subworkflows, and approval gates.
- Worked with the Studio information architecture across workflows, deployments, runs, schedules, files, knowledge, user tables, integrations, dashboards, and chat-based interaction.
- Evaluated whether operator flows exposed enough context to review an action before approving it and enough run state to diagnose what happened afterward.
- Helped position finance recipes as implementations on top of reusable platform primitives rather than hard-coding the product as a single-purpose finance assistant.
- Tested access-controlled demo paths and translated operational feedback into product and workflow requirements during the nine-month cycle.
Contribution summary
- Founded Agent Mag and leads the roadmap across the publication, builder ecosystem, agent experience, and operations automation product.
- Implemented Azure OpenAI GPT-4o streaming, then built a multi-step function-calling loop with visual tool-result blocks for production agent interactions.
- Integrated an OpenClaw gateway and model routing, then hardened the chat surface with identity lock, security rules, persona controls, and an explicit OpenClaw workspace.
- Translated finance and operations procedures into typed workflow stages, approval points, integration boundaries, and operator-visible execution state.
- Uses Claude Code and Codex to move from product requirements through implementation, debugging, testing, deployment, and operational follow-through.
Failure design
Reliability engineering
R01
Risk
A long workflow outlives a browser request or worker process.
Control
Durable Trigger.dev execution and persisted run state move work outside the request lifecycle and make progress observable.
R02
Risk
A model proposes an unsafe or irreversible action.
Control
Typed action blocks and explicit approval nodes separate model reasoning from execution authority.
R03
Risk
An operator cannot explain a historical outcome.
Control
Run outputs, errors, timestamps, approval decisions, and audit events are stored against the workflow and company context.
R04
Risk
Provider credentials leak into browser state or prompts.
Control
Connections are handled through the server-side integration layer rather than embedded in client workflows.
R05
Risk
Work from one company is applied to another.
Control
Company and workspace context is carried through operational records and execution boundaries.
Tradeoffs
Technical decisions
Separate agents from workflows
Agents are useful where interpretation is needed; deterministic workflows remain easier to review, retry, and govern for consequential operations.
Treat approvals as runtime primitives
Human review must survive refreshes and worker restarts, so it belongs in persisted execution state rather than an informal chat step.
Build domain recipes on reusable modules
The same email, document, CRM, accounting, table, wait, and approval blocks can support more than finance without duplicating the runtime.
Credibility
Evidence and scope
Founder-owned GitHub organization
Harshitha is the sole public organization member and active owner of Agent-mag, which contains the product, skills registry, tools, documentation, roadmap, and community repositories.
Direct implementation history
Her product commits cover Azure OpenAI streaming, a function-calling chat loop, visual workflow blocks, OpenClaw gateway integration, model routing, security and identity controls, persona workspace files, and interface refinement.
Reviewed private codebase
The current monorepo contains four deployable applications, 16 core workflow blocks, 17 provider integrations, a PostgreSQL schema with 31 migrations, and extensive run, approval, file, knowledge, table, and workflow surfaces.
Access-controlled product demo
The private Cloudflare demo exposes the current operator product while protecting internal access.
Attribution boundary
Harshitha is Agent Mag's founder, product owner, and active organization owner. Her direct Git history verifies the named agent-chat, OpenClaw, security, persona, and interface implementations. The platform counts are a reviewed codebase snapshot; the finance-runtime architecture describes the wider product she has led and operated across its nine-month build cycle.
Technology