Agent System Overview
Forge isn’t a single AI. It’s a network of specialized agents, each built for a specific task. This is what makes it feel like you're chatting with an on-chain intelligence that knows exactly what it's doing.
This section explains what Forge agents are, how they’re structured, and why they’re the backbone of the entire system.
🤖 What Is a Forge Agent?
A Forge agent is a modular AI process designed to handle one specific type of blockchain intelligence task. Each agent has access to the Solana indexer, listens for key events, and can take context from MCP to return tailored insights.
Think of them like:
One agent for watching wallets
One for monitoring liquidity pools
One for understanding token deployers
One for reading Telegram sniper movement
Instead of trying to make one AI do everything, Forge spins up the right specialist — and lets that agent operate in its own optimized logic environment.
🏗 Agent Structure
Each Forge agent has 5 core components:
Purpose – What it’s designed to do
Context Hooks – What live data streams it listens to
Execution Logic – How it processes information
Prompt Template – What kind of answers it gives
Response Format – How it replies to users or other agents
Let’s break those down:
1. Purpose
Every agent is goal-oriented. It’s created with a specific use case in mind:
TokenAgent → detect trends and anomalies in new tokens
WalletAgent → track buy/sell activity, identify farming patterns
LPAgent → monitor burn events, LP locks, tax changes
SniperAgent → watch sniper buys, trace click-to-buy wallets
TelegramAgent → detect KOLs joining new chats, identify signals
If your question aligns with the agent’s purpose, it will automatically be triggered.
2. Context Hooks
Agents pull their inputs from live streams of Solana blockchain data. These are the real-time hooks that feed them fresh context.
Hooks can include:
“New Token Created”
“LP Added or Burned”
“Large Wallet Buy Over X SOL”
“Telegram Join By Known Wallet”
“Creator Renounced Contract Ownership”
Agents listen for these constantly, and some will only wake up when their hook conditions are met.
3. Execution Logic
This is the agent’s internal decision tree — how it filters data, evaluates risk, scores relevance, and decides what to say.
Examples:
A WalletAgent might tag a wallet as a sniper if it buys 3+ tokens within 10 seconds of creation
A TokenAgent might flag a deployer if they’ve launched 5 rugs in 24 hours
An LPAgent might evaluate liquidity depth vs. market cap and issue a risk warning
This logic is customizable, meaning Forge devs (or users with advanced access) can tweak thresholds and behaviors.
4. Prompt Template
Each agent has a custom-written prompt structure that shapes how it speaks.
Some are formal and statistical:
“This wallet has interacted with 9 tokens in the last 3 hours, showing a clear pattern of early entry and early exit behavior.”
Others are designed to be more casual or alert-style:
“🚨 This sniper is back. Just entered a fresh meme with 30 SOL. Last seen in 3x rugs.”
The tone can be tuned depending on audience (devs vs. traders vs. analysts).
5. Response Format
Agents can:
Return a simple answer in chat
Open a modal with charts
Trigger a follow-up question
Send context to another agent for collaboration
Push on-chain memos or flags
Forge is not just a passive assistant. Agents can trigger on-chain activity, alert teams, or even influence decision-making pipelines.
🔁 Multi-Agent Collaboration
Sometimes, a single question needs multiple agents. For example:
“Is this token safe and who deployed it?”
Forge will:
Trigger TokenAgent → to scan the contract
Trigger WalletAgent → to check the deployer’s behavior
Trigger LPAgent → to verify liquidity structure
Each agent returns its part, then Forge compiles the response as if it’s a single smart analyst.
This is where Forge becomes more than a dashboard — it becomes collaborative intelligence.
🧠 Memory and State
Most agents are stateless, meaning they analyze based on real-time data without remembering past interactions. However, some advanced agents maintain short-term memory for deeper analysis chains, like:
“What has this wallet done since I last asked?”
“Did this token reach my alert threshold today?”
State memory is kept lightweight for performance and privacy.
🔧 Developer Notes
Advanced users and Forge integrators can:
Build their own agents using MCP
Override the default prompt template
Add custom event hooks
Push data from external sources (e.g. Twitter, Telegram, TradingView)
Forge isn’t closed. It’s a framework for AI-agent logic on-chain.
Last updated