Fetch Agent Mesh

Build Agents

AI Builder Agent

Tell the builder what you want in plain English. It creates agents, skills, connectors, edits code, and promotes to the mesh — all through conversation.

Create & Compose

Scaffold agents from skills and system prompts. Compose freely — tools are discovered at runtime, no hardcoded dependencies between agents.

Skills & Connectors

Skills are prompt instructions. Connectors are deterministic tools (HTTP calls, data processing). Agents discover and use them at runtime.

Any AI Model

AWS Bedrock, OpenAI, or Anthropic — Claude, GPT, Llama, MiniMax, o3. Pick per agent, switch without code changes. LiteLLM routes automatically.

Bring Any Agent

Add existing agents with zero changes to their code. External APIs, third-party services, microservices — wrap as a connector and they’re discoverable.

Dynamic Discovery

Agents find each other on the mesh by description and tags. No routing config — the LLM decides what to call based on what’s available.

Deploy & Operate

Personal, Team & Company Meshes

Create personal meshes for individual users, team meshes for domains, and company-wide meshes — all without duplicating agents. Simple namespace isolation means the same agent code works at every scope.

One-Command Deploy

mesh promote my-agent — validates, versions, snapshots, copies to the mesh, and auto-starts. Rollback to any previous version instantly.

5 Interfaces

Web UI, Terminal TUI, bash CLI, REST API, and Claude Code skills. Every operation works from anywhere — pick the interface that fits your workflow.

Cross-Agent Observability

Full call chain tracking with real-time visualization. See which agents are active, what they’re calling, and how long each step takes.

Secrets & Config

Environment-aware config per connector. Secrets from AWS Secrets Manager. No env vars — just resolve_connector_config("my-tool").

Agent Identity

DID cryptographic identity for every agent. Ed25519 signed inter-agent calls. Verifiable Credentials for tamper-proof execution audit trails.

Elastic K8s Runtime

Ephemeral agent pods scale to zero when idle and spin up on demand. Per-user agent pools, scheduled cron jobs, and idle scanning keep costs minimal while staying responsive.

---

Bring Any Agent Into the FAM

FAM doesn’t require agents to be built with FAM. Any existing service can join the mesh:

External APIs — Wrap any HTTP API as a connector. The Fetch consumer agent, FAST analytics, offer programming — all are external services that mesh agents call through connectors with zero changes to the underlying service.

Third-party agents — If it speaks HTTP, it can be a connector. Slack bots, Jira integrations, GitHub APIs, LangChain agents, custom microservices — wrap the endpoint and it’s discoverable.

How it works:

1. Create a connector with config.yaml pointing to the service URL
2. Write a thin impl.py that calls the API
3. Create a skill that tells agents how to use it
4. Every agent on the mesh can now discover and call it

# Example: add an existing analytics API
mesh tools create analytics-api --function-name query_analytics
# Edit tools/analytics-api/config.yaml with the URL per environment
# Edit tools/analytics-api/impl.py with the HTTP call
mesh skills create analytics-query --tool analytics-api
mesh create analytics-agent --skill analytics-query
mesh promote analytics-agent
# Done — every agent on the mesh can now query your analytics API

No changes to the analytics API. No SDK to install. No protocol to implement.

---

The Builder Agent

FAM includes an AI-powered builder agent that can create and manage your entire mesh through conversation:

“Create an agent that queries Fetch analytics data about brand performance”

The builder will:

• Select appropriate skills from the library
• Generate a system prompt tailored to the task
• Scaffold the agent with proper tool discovery
• Create any missing connectors with real implementation code
• Test and promote it to the mesh

Available in the Web UI (right panel), TUI (mesh builder), and via the REST API (POST /chat/stream).

---

Use It From Anywhere

Every mesh operation is accessible through 5 interfaces:

Operation	CLI	Web UI	TUI	Claude Code	API
Create agent	mesh create	Builder chat	Builder chat	/scaffolding-mesh-agents	POST /agents
Start / Stop	mesh start / stop	Start button	start_dev tool	/executing-mesh-agents	POST /agents/{n}/dev
Test	mesh test	Test button	test_agent tool	/testing-mesh-agents	POST /agents/{n}/test
Promote	mesh promote	Promote button	promote_agent tool	/promoting-mesh-agents	POST /agents/{n}/promote
Execute	mesh exec	Agent chat	Agent chat	/executing-mesh-agents	POST /execute
Interactive chat	mesh chat	Agent chat	Agent chat	/executing-mesh-agents	POST /execute/stream
Mesh status	mesh status	Live indicators	—	/executing-mesh-agents	GET /mesh/nodes
Manage skills	mesh skills	Skill panel	Builder	/managing-mesh-agents	GET/PUT /skills
Manage connectors	mesh tools	Connector editor	Builder	/managing-connectors	GET/PUT /connectors

---

Deploy in 60 Seconds

Quick Start

# Install
git clone https://github.com/fetch-rewards/labs-inno-agent-platform-poc-2
cd agent-mesh && python -m venv .venv && source .venv/bin/activate
pip install -e .

# Start the mesh
af server --port 8090 --no-vc-execution --open=false &
mesh server &

# Pick your interface
mesh builder              # Terminal UI
cd web && npm run dev     # Web UI at localhost:5173
mesh chat consumer-agent  # CLI chat

Create an Agent

# From the CLI
mesh create my-agent \
  --skill fast-analytics-query \
  -d "Answers analytics questions" \
  --model claude-sonnet-4-6

# Start it
mesh start my-agent

# Talk to it
mesh exec my-agent.think "top soda brands last week"

Promote to Mesh

# Validate
mesh test my-agent

# Preview what will happen
mesh promote my-agent --dry-run

# Deploy — bumps version, snapshots, auto-starts
mesh promote my-agent

# Verify
mesh status

---

Agents on the Mesh

Agent	What it does	Status
consumer-agent	Bridges to Fetch Rewards consumer API for customer support	Promoted
fast-analytics-agent	Queries FAST for brand, category, retailer analytics	Promoted
customer-agent	Simulates a real customer interacting with mesh services	Promoted
account-lead	Manages client relationships and programs offers	Promoted
github-collector	Analyzes GitHub repos and PRs	Active
analyst	Scores, ranks, and generates digest reports	Active
researcher	Searches sources and synthesizes findings	Active
architect	Analyzes system architecture and proposes ADRs	Active

---

Architecture at a Glance

┌─────────────────────────────────────────────────────────┐
│                AgentField Control Plane                  │
│     Registration · Discovery · Execution · Memory        │
└────────┬──────────────┬──────────────┬──────────────┬───┘
         │              │              │              │
   ┌─────┴─────┐  ┌─────┴─────┐  ┌─────┴─────┐  ┌───┴───┐
   │ consumer  │  │   fast    │  │ customer  │  │  ...  │
   │  agent    │  │ analytics │  │  agent    │  │agents │
   └─────┬─────┘  └─────┬─────┘  └───────────┘  └───────┘
         │              │
   ┌─────┴─────┐  ┌─────┴─────┐
   │ Stage     │  │ Stage     │
   │Consumer AP│  │ FAST API  │
   └───────────┘  └───────────┘

Interfaces:
  Web UI (React) · TUI (Textual) · CLI (Typer) · Claude Code · REST API