Loading…
Loading…
Complete guide to using AGENCY BAZAAR — your agentic organization platform for building, deploying, and managing AI-powered multi-agent workflows.
AGENCY BAZAAR is an agentic organization platform that lets you build intelligent multi-agent systems without writing code. Think of it as a visual operating system for AI — you design teams of AI agents, connect them with tools, choose how they coordinate, and launch workflows that run autonomously on your behalf.
This guide covers every feature available in the current release. Whether you are setting up your first agency from a template or building a custom pipeline with parallel agents and human approval gates, this document walks you through it step by step.
Alpha Release
Version 0.6.0 is the latest alpha. This release delivers production-grade architecture: per-service PostgreSQL split with PgBouncer, CSRF/CSP edge enforcement, account lockout, cost attribution, latency profiler, adversarial testing, OWASP ZAP CI, quota enforcement, blueprint version diff, webhook replay, workflow checkpoint replay, transactional outbox, W3C distributed tracing, and full i18n namespace splitting. All 18 pages are decomposed with TanStack Query hooks. Expect continued rapid iteration.
You don't need an account to start exploring. Unauthenticated visitors see a Guest mode: read-only badge in the header and can freely browse Templates, Marketplace, Tools, Workflows, the User Guide, and What's New.
Write actions (creating agencies, installing blueprints, running workflows, registering tools) will display a friendly alert prompting you to create an account first.
Templates are pre-built organizational blueprints that represent real-world business units. Each template bundles a ready-to-use set of AI agents, tool configurations, and an orchestration pattern — so you can start automating in seconds.
Citizen services, permit processing, public records management
Enrollment workflows, faculty coordination, course approvals
Patient intake, scheduling, compliance reporting
Engineering sprints, product research, growth analytics
Risk analysis, compliance checks, client reporting
Inventory management, demand forecasting, supplier comms
Production scheduling, quality control, procurement
Use the All / Public / Private tabs above the template grid to narrow results. Your selection persists while you browse.
Saved blueprints maintain version history. Open a blueprint detail page to compare two versions side by side — added agents are highlighted in green, removed agents in red, changed configuration in amber, and unchanged nodes in grey. This makes it much easier to review template evolution before you publish updates to the Marketplace or reuse a blueprint in a new workflow.
The Workflow Builder is a visual canvas where you design multi-agent pipelines. Navigate to Workflows → New Workflow or click Buildon an existing workflow to open it.
Use the Pattern Selector panel on the right to choose how agents cooperate. See the Orchestration Patterns section for a full description of each option.
Before saving to the backend, the builder runs a validation pass. Warnings are shown for common issues such as disconnected agents, missing system prompts, or an unsupported pattern for the number of agents configured.
Orchestration patterns define how your agents communicate and hand off work. Choose the right pattern for your use case to get the best results.
Agents execute one after another in a fixed chain — A finishes, then B starts, then C. The output of each agent becomes the input to the next.
Best for: pipelines with dependent steps, e.g. Research → Draft → Review → Publish.
All agents run at the same time (fan-out). Their outputs are collected and merged into a single aggregated result (fan-in).
Best for: tasks that can be split across agents independently, e.g. analysing multiple documents simultaneously.
A supervisor agent receives the task, breaks it down, delegates sub-tasks to worker agents, then assembles the final result.
Best for: complex tasks requiring planning and delegation, e.g. a manager coordinating a cross-functional team.
A router agent analyses the incoming request and directs it to the most appropriate specialist agent based on intent or topic.
Best for: support desks or multi-department systems where different queries need different handlers.
An agent iterates repeatedly on its output, refining it each cycle until a quality threshold or maximum iteration count is reached.
Best for: creative writing, code generation, or any task that benefits from self-critique and improvement.
The workflow pauses at defined checkpoints and waits for a human to review, approve, or modify the agent's output before continuing.
Best for: high-stakes decisions, compliance reviews, or any step requiring human judgement.
Agents communicate peer-to-peer in a dynamic mesh. An LLM-powered router selects the next speaker each round, enabling emergent coordination without a central controller. Configurable max_rounds safety cap prevents runaway conversations.
Best for: brainstorming, debates, or exploratory tasks where the optimal agent sequence is not known upfront.
All agents run in parallel (fan-out), then a dedicated LLM-powered synthesiser merges, deduplicates, and resolves contradictions across their outputs into a single unified result.
Best for: research synthesis, multi-perspective analysis, or any task where diverse viewpoints need to be consolidated.
The Marketplace is a community hub where users publish, discover, and install pre-built agent blueprints. Find a blueprint that fits your needs, install it in one click, and run it immediately.
Visit Marketplace → Analytics to see platform-wide statistics: top categories by install volume, most active creators, aggregate blueprint counts, and trending patterns.
Tools extend what your agents can do. A tool is a capability — web search, code execution, sending email, querying a database — that an agent can call during workflow execution.
Click Install on a tool card to add it to your agent's available tool set. Installed tools are immediately available for selection inside the Workflow Builder's agent configuration panel.
Workflows are live instances of a blueprint — they hold execution state, track progress, and accumulate results over time.
The monitor screen auto-refreshes every 2 seconds and displays:
Once a workflow reaches COMPLETED status, click View Results to open the results page. You will see the full agent output, structured by step, along with any files or data generated during execution.
If a workflow fails partway through, you can resume it from any completed step rather than starting over. Open the failed workflow and click Replay. Select the step index to restart from — the platform skips already-completed agents and continues from that checkpoint, preserving all prior agent outputs.
Workflow runs can be constrained by both per-run budgets and your plan's monthly token allowance. If a workflow would exceed its configured cost or token budget, the run stops early. If your tenant exhausts its plan quota, new runs are rejected with a quota error until the next monthly reset or a plan upgrade.
The Memory Service gives your agents persistent, context-aware recall across sessions. Instead of starting fresh every run, agents can remember facts, past conversations, and learned strategies.
Current conversation context. Automatically expires after the session or TTL period.
Durable facts and user preferences that survive across all sessions.
Memories of past conversations, stored as semantic embeddings for fuzzy recall.
Structured knowledge base and domain concepts, searchable by meaning.
Learned patterns and strategies — how the agent has found best to approach a task.
Memory is managed automatically during workflow execution. You can also manually store or retrieve memory entries via the API for advanced use cases — for example, pre-loading domain knowledge before a workflow runs or extracting summaries after completion.
The platform automatically captures execution telemetry for every workflow and agent run. Use the analytics dashboard to understand performance, cost, and usage patterns.
Visit Marketplace → Analytics for platform-wide insights: top blueprint categories, most active creators, total installs, and usage trends over time.
Use the Status page for a quick platform health snapshot, then move into the observability dashboards when you need more depth. The latest service-health views bring together availability, error budget, latency, active alerts, circuit-breaker state, saga compensation, and downstream failure clues in one place so you can tell whether an issue is isolated or platform-wide.
The platform tracks per-execution LLM spend and rolls it up by tenant. The Cost card on your dashboard shows total USD spend for the current billing period. Detailed breakdowns are available at GET /api/v1/telemetry/cost/{tenantId}.
The Observability page includes a Latency Profile card visualising P50, P95, P99, and P99.9 percentiles per agent over a 30-day sliding window. Use this to identify slow agents or capacity bottlenecks before they affect users.
Platform administrators can export all traces to Google BigQuery for deeper analysis using SQL. This requires GOOGLE_CLOUD_PROJECT and BIGQUERY_DATASET configuration. See your administrator for access.
All user-supplied text passed to agents is scanned for prompt injection patterns before reaching the LLM. Suspicious inputs are blocked with an error rather than silently passed through.
Agent outputs are automatically scanned and redacted for personally identifiable information including Social Security Numbers, credit card numbers, email addresses, and phone numbers before the results are stored or displayed.
All data-mutation API routes (POST, PUT, PATCH, DELETE) are protected against cross-site request forgery via Origin-header validation in the edge middleware. A strict Content-Security-Policy is enforced platform-wide to block inline script injection, alongside X-Frame-Options, Referrer-Policy, and Permissions-Policy headers on every response.
Generate long-lived personal API tokens in Settings → API Tokens for programmatic or CI/CD access. Tokens use the ab_pat_ prefix and are accepted in place of a JWT Bearer token on all endpoints.
All your blueprints, workflows, memory entries, and tools are scoped to your tenant account. No other user or organisation can see or access your data.
AGENCY BAZAAR supports guest access so that anyone can explore the platform without creating an account. This is ideal for evaluating features before signing up.
When you visit the platform without logging in, the header displays a Guest mode: read-only badge instead of user controls. If you attempt a write action, a friendly alert appears explaining you need an account and offering a link to sign up.
AGENCY BAZAAR supports dark and light colour themes so you can work comfortably in any environment. Your preference is remembered across sessions.
Clean, bright interface optimised for well-lit environments. This is the default for first-time visitors.
Low-contrast interface that reduces eye strain in dim environments and saves battery on OLED displays.
Automatically matches your operating system preference. Switches in real time if you change your OS theme.
prefers-color-scheme media query.On your first visit each browser session, a cinematic splash screen plays — showing the Perpetual Works studio card, followed by the AGENCY BAZAAR logo with the tagline "Build the future with AI agents". The animation lasts about 5 seconds and can be skipped at any time using the Skip button. The splash screen only plays once per session.
AGENCY BAZAAR is deployed to Google Cloud Run for production hosting. All 9 services (8 backend microservices + Next.js frontend) are containerised and managed via Terraform.
.env with GCP project ID and API keys.gcloud auth login and gcloud auth configure-docker.terraform init && terraform apply in infrastructure/terraform/.For local development, use docker-compose up to start infrastructure (PostgreSQL, Redis, Kafka, MinIO), then run individual services or use the scripts/start_all.ps1 PowerShell script to launch all 8 backend services with correct port bindings.
Every backend service exposes standard health probes, and production rollout automation now validates live metrics before fully promoting new revisions. During a staged canary rollout, the pipeline checks live error-rate, latency, readiness-burn, and restart signals before moving from 5% to 50% to 100% traffic. If the signals degrade, the rollout is designed to stop and roll back rather than continue blindly.
AGENCY BAZAAR User Guide · Version 0.4.0 Alpha · Last updated February 2026
This document is updated with every new release. Check What's New for the latest changes.