virattt/dexter
An autonomous agent for deep financial research
A practical agent workflow for finding signal before CT finds exit liquidity.
Turn what you learned into a concrete stack decision.
Want the shortlist in your inbox?
Subscribe for the weekly brief that turns new AI noise into the few tools and workflows worth testing.
An autonomous agent for deep financial research
Guide
AI Repos Weekly: July 31, 5 Tools Actually Worth Your Attention
Marketing agencies are getting quietly automated — here's the stack doing it
Guide
Why a Research Agent Beats a Trading Bot for DeFi DD
Autonomous trading bots run on vibes. This one runs on filings.
Guide
Dexter: The Autonomous Research Agent Crypto Diligence Actually Needs
An open-source agent that pulls financials, cross-checks sources, and ships a report while you do something else.
TradingAgents: Multi-Agents LLM Financial Trading Framework
Crypto research used to mean 47 tabs, three half-read docs pages, a Discord search, a Dune dashboard, and one suspiciously confident thread from a guy with a Pudgy Penguin PFP.
That still works if you have unlimited time and a high tolerance for nonsense.
I use agents differently. Not as “press button, get alpha” machines. That is mostly fantasy. I use them as research interns that can read faster than me, argue with each other, and surface things I would otherwise miss.
My current stack is simple:
The important part is the order. Agents first for breadth. Humans last for judgment.
AI agents are good at pattern finding.
They can scan docs, summaries, market data, governance posts, protocol announcements, and competitor pages faster than you can. They are useful for weak signals: repeated complaints, weird incentive shifts, user growth that does not match token price, vague roadmap language, or a new integration that has not hit Crypto Twitter yet.
They are bad at truth.
They do not know when a founder is massaging metrics. They do not know when CT is quietly coordinated. They do not feel the difference between real demand and mercenary liquidity unless you force them to check.
So I do not ask agents: “Should I buy this token?”
That is a bad question.
I ask:
“What would have to be true for this to be a good trade, and what evidence supports or breaks that thesis?”
That framing changes everything.
I use Dexter when I want a structured research brief.
Good use cases:
Dexter is best when you give it a narrow target and force it to cite evidence. If you ask for “research Solana DeFi,” you get soup. If you ask for “compare Kamino, MarginFi, and Drift on lending growth, revenue quality, and token capture,” you get something useful.
I use TradingAgents after I already have a rough view.
Its value is debate. I want a bull case, bear case, risk manager view, and final synthesis. Not because the output is magically correct, but because it prevents me from falling in love with my first idea.
This matters in crypto because most bad trades start as good stories.
“Real yield.” “AI agent infra.” “Restaking ecosystem.” “Modular liquidity.” “Next Hyperliquid.”
All of those can be valid. All of them can also be exit liquidity with better typography.
Here is the research loop I use before I take a token seriously.
Bad prompt:
“Research this token.”
Better prompt:
“Research whether TOKEN is mispriced relative to its fundamentals over the next 3-6 months. Focus on revenue quality, user growth, token unlocks, liquidity depth, competitor positioning, and upcoming catalysts.”
That prompt gives the agent a job. It also defines the time horizon. A token can be a bad long-term investment and still a strong 30-day trade. Mixing those is how people confuse themselves.
My Dexter prompt usually looks like this:
Create a crypto research memo for [PROJECT/TOKEN].
Goal: decide whether this deserves deeper manual research, not whether to buy.
Cover:
1. What the protocol does in plain English
2. Core users and why they use it
3. Revenue, fees, TVL, volume, active users, or closest available metrics
4. Token utility and value capture
5. Token supply, emissions, unlock schedule, and major holders
6. Main competitors
7. Upcoming catalysts in the next 3-6 months
8. Biggest risks and what would invalidate the bull case
Rules:
- Separate facts from assumptions
- Flag missing data
- Do not use CT sentiment as evidence unless tied to a measurable event
- Give me a final score from 1-10 for “worth deeper research”
The final score is not the point. The missing-data section is the point.
If Dexter says “token unlock data unavailable” or “revenue source unclear,” that is where I go manually. Most people do the opposite: they read the confident parts and ignore the gaps.
Once I have a memo, I feed the thesis into TradingAgents.
Example prompt:
Analyze this crypto trade thesis using multiple agents.
Thesis:
[TOKEN] may outperform over the next 90 days because [catalyst], [metric trend], and [relative valuation].
Assign roles:
- Bull analyst
- Bear analyst
- Market structure analyst
- Risk manager
- Final portfolio manager
Each role should:
- State its strongest argument
- State what evidence would change its mind
- Identify time horizon mismatch
- Identify reflexive risks from CT narratives
Final output:
1. Best version of the bull case
2. Best version of the bear case
3. What must be checked manually before acting
4. Position sizing view: avoid, watchlist, small, medium, or high conviction
I care most about the risk manager. If the risk manager gives me generic nonsense like “crypto is volatile,” the setup is not ready. I want specific risks: cliff unlock in 21 days, TVL subsidized by points, three wallets controlling governance, perps OI too crowded, or FDV already pricing perfect execution.
Let’s say I am looking at a mid-cap DeFi token trading at:
I am not naming a token here because the point is the process, not a stale call.
First, I ask Dexter:
Research this setup as if I am considering a 30-90 day trade.
Known inputs:
- Token price: $2.40
- Circulating market cap: $240M
- FDV: $1.2B
- Float: 20%
- 30-day protocol fees: $3M
- v2 launch expected in six weeks
- 8% supply unlock in 35 days
Find:
1. Whether fees are organic or incentive-driven
2. Whether v2 has confirmed launch details or just marketing
3. Token value capture from protocol fees
4. Top competitors and their valuation multiples
5. Any governance votes, emissions changes, or community issues
6. Whether the unlock is likely material relative to daily volume
Then I calculate a few simple numbers myself.
Annualized fees:
$3M monthly fees x 12 = $36M annualized fees
FDV / annualized fees = $1.2B / $36M = 33.3x
Circulating market cap / annualized fees = $240M / $36M = 6.7x
That looks cheap on circulating market cap and less cheap on FDV. This is where beginners get trapped. If unlocks are heavy, FDV matters more than the chart wants you to believe.
Now I check unlock pressure:
Total supply implied by FDV:
$1.2B / $2.40 = 500M tokens
Unlock amount:
500M x 8% = 40M tokens
Unlock value:
40M x $2.40 = $96M
If daily spot volume is only $12 million, that unlock is eight days of full spot volume. Not all unlock recipients dump, but you do not ignore $96 million of potential supply because the Discord says “long-term aligned.”
Next, I send the whole setup into TradingAgents:
Debate this DeFi token trade.
Facts:
- Price $2.40
- Market cap $240M
- FDV $1.2B
- Annualized fees around $36M
- FDV/fees around 33x
- Market cap/fees around 6.7x
- v2 launch expected in six weeks
- 8% supply unlock in 35 days, about $96M at current price
- Daily volume appears materially lower than unlock value
Question:
Is this a good 30-90 day trade, or is the unlock likely to cap upside?
Give bull, bear, market structure, and risk manager views.
The output I want is not “buy” or “sell.”
A useful answer looks like:
My likely decision: watchlist, not buy.
If price sells off after the unlock and fees keep growing, then I revisit. That is the edge. Not predicting perfectly. Avoiding obvious bad timing.
| Approach | Best for | Weakness | |---|---|---| | Manual tab research | Final verification, judgment, tokenomics | Slow and easy to bias toward your first narrative | | Dexter | Fast project memo and weak-signal discovery | Can miss context or overtrust available sources | | TradingAgents | Stress-testing a thesis | Only as good as the facts you feed it | | CT threads | Narrative temperature | Often polluted by bags, engagement farming, and paid promotion | | Paid research groups | Curated deal flow | Quality varies wildly, and incentives are not always clean |
If I had to pick only one, I would use Dexter plus manual checks.
TradingAgents becomes more valuable once you are already tempted to buy. It is basically a structured argument machine. That is useful because the moment you want the trade to work, your brain starts filtering evidence.
I skip agents for pure momentum trades.
If a token is moving because Binance listed it, a major perp market opened, or a short squeeze is underway, I do not need a 12-page memo. I need market structure, liquidity, and risk limits.
I also skip agents for obvious vaporware. If the website is vague, the docs are empty, the token is already at a silly FDV, and the only catalyst is “AI narrative,” I do not need to outsource common sense.
Agents are for situations where there is enough substance to analyze.
Do not ask, “Where will this token be next month?”
Ask what would drive repricing, what could block it, and what data confirms or breaks the thesis. Price targets are usually fake precision.
Crypto Twitter is useful for narrative timing. It is not proof.
If an agent says sentiment is positive, ask why. Is volume rising? Are new users showing up? Are fees increasing? Did a real integration ship? If not, it may just be coordinated posting.
Low float makes charts look better than fundamentals.
Always check circulating supply, FDV, unlock dates, emission rate, and who receives the unlock. A “cheap” token can be expensive once future supply enters the room.
Agents can lag. Even when they browse or ingest fresh sources, they can still miss the latest governance vote, exploit rumor, exchange listing, or token unlock update.
Before acting, manually check the project docs, token unlock calendar, official X account, governance forum, and largest venues where the token trades.
I use agents to answer three questions:
That last question matters most.
The edge is not “AI finds alpha.” That is too vague. The edge is compressing the boring part of research so you have more time for the hard part: judgment.
Crypto rewards speed, but it punishes lazy certainty. Agents help with speed. They do not fix lazy certainty.
Yes, if you use them for summarizing, comparing, and stress-testing. They are not reliable enough to make buy or sell decisions without manual checks on tokenomics, unlocks, liquidity, and source quality.
For my workflow, Dexter is better for deep research memos, while TradingAgents is better for debating a trade thesis. They solve different problems.
Sometimes they can surface weak signals early, like unusual growth, ignored catalysts, or risk factors people are not discussing yet. But they also repeat bad data confidently, so the alpha comes from combining agent speed with manual verification.
→ Ask the index what to build your web3 ai stack
→ Free credits for these tools
Written by McKlaud AI. Want to know which AI tools actually fit your business? Get a free AI audit.