Letting a team of AI agents manage my crypto portfolio

I'm the founder and creator of the AI agent platform Nelly, and while experimenting with a Crypto Analyst agent, I got curious about its real-world performance. So I decided to let a team of AI agents create and manage a crypto asset portfolio for me, actually three portfolios: one conservative, one balanced, and one aggressive. The agents will choose which crypto assets to research, analyse them one-by-one with a 6-month outlook, and then assemble portfolios that match each risk level based on those reports. I'll have the agents manage these portfolios monthly and track their performance against a few benchmarks.

Note: These aren't day-trading bots chasing patterns. They're investment agents that construct a portfolio from their own research and then monitor it periodically.

The AI agent platform

Nelly is a platform for creating AI assistants using natural language. Agents in Nelly are built by breaking a problem into specialized sub-agents, each with its own set of instructions. They run locally on your computer and can use tools such as a browser, a calculator, and a JavaScript runtime. The LLM used is Anthropic's Claude 4 Sonnet, and I used ChatGPT to help draft the specs for the agents in this article. I've published them to the Nelly Marketplace so you can see how they work, try them yourself, or even improve them and make your own versions.

Scouting for crypto assets

To analyze crypto assets, the agents first need a shortlist. For this I created the Crypto Token Scout agent, which compiles a diversified list of tokens based on constraints like "exclude stablecoins", minimum 24-hour volume, and more.

Crypto Token Scout · AI Agent · Nelly

Get a diversified list of crypto assets to research for your portfolio. AI agent by timothy.

I then asked it for 20 good candidates:

Suggest 20 tokens that are good candidates to research further for inclusion into a 6-month portfolio.

And in under 10 minutes it came back with these 20 crypto tokens:

Read the entire output with methodology here: https://timothyej.com/crypto-token-research-queue-20-selected-assets/

Reviewing the list

The list contains 20 tokens as requested and they all have a minimum 24h volume of $5M, are at least 30 days old, and listed on at least 3 exchanges. The diversity caps are also respected (max share per sector: 35%, max share per network: 40%) and the list has no stablecoins.

While the list is correct, the summary has a couple of errors:

• ⚠️ There are 6 Ethereum Network tokens, not 5.
• ⚠️ The "data fetched" timestamp should be 13 August 2025, not "2025-01-13"

Conclusion: These are minor mistakes and in this case they don't matter as the summary is not being used down stream.

Researching and analyzing the crypto assets

With a diverse list of crypto assets, I built the Crypto Analyst agent to research and analyze each token and output a comprehensive, decision-ready brief for each:

Crypto Analyst · AI Agent · Nelly

Get a comprehensive crypto analysis report. AI agent by timothy.

I had the agent analyze each candidate one-by-one using this prompt:

Analyze [Name]'s ([<Symbol]) performance and outlook for the next 6 months, including technical analysis and market sentiment.

On my machine it took about 15-30 minutes per asset. I saved all reports as Markdown files for the next step. Nelly agents can only run one task at a time, so I spread this over a couple of days (total: 6h 47m 21s). You can find all reports below, unedited (except for the main title):

Costs

To produce all of these reports cost me ~$130 in LLM costs.

Reviewing the reports

It would have been too much work to verify every report, so I picked 5 of them. I checked both manually and with the help of ChatGPT, and verified all of ChatGPT's claims before including them here. Treat this as a sanity check rather than a full audit.

Note: I did not fix any of the mistakes AI did as that would somewhat go against the idea of having AI manage my portfolios.

AAVE

• "$35.61B in Total Value Locked (TVL)" ⚠️ Correct according to CoinMarketCap, but conflicting with Token Terminal ($58B).
• "Active Users: +5.8% (94,500 unique addresses)" ⚠️ Seems to be +16.5% (https://tokenterminal.com/explorer/projects/aave/metrics/)

LTC

• "Lightning Network: Planned integration for microtransactions" ⚠️ Litecoin already has Lightning Network.
• "ETF Confidence: "All-time high" confidence from industry experts" ⚠️ Based on old news, Litecoin ETF decision delayed to October 2025.
• "Sharpe Ratio: 2.35 (excellent risk-adjusted returns)" ⚠️ Looks like this was calculated based on simulated historical price data.

ETH

• "significant 46.5% recovery from July's low of $2,942" ⚠️ July's low was even lower than that, near $2,400.
• "Pectra Upgrade: Scheduled for May 2025, expected to enhance scalability and efficiency" ⚠️ Stale news.

SHIB

• "Shibarium L2: Over 715M transactions processed, growing ecosystem" ⚠️ Now at over 1.5B transactions.
• "Surpassed 1M holders milestone in July 2025" ⚠️ Seems like it surpassed 1M holders much earlier than July 2025.

LINK

• "Rally Performance: +125% from $11.65 June low" ⚠️ That would a 116% increase, not 125%.
• "vs Bitcoin: LINK +43.6% vs BTC +8.2% (30-day)" ⚠️ BTC was down almost 3% (30-day).
• "vs Ethereum: LINK +43.6% vs ETH +12.8% (30-day)" ⚠️ ETH was up between 18-25% (30-day).
• "LINK Reserve Program Launch: July 2025" ⚠️ This was announced on Aug 7.

Conclusion: Based on the reports I reviewed, the outputs are accurate enough for this experiment. However, I do have some ideas on how to improve their accuracy further, but that's for the next iteration.

Allocating the portfolios

With 20 candidates and fresh 6-month analyses, I built the Crypto Portfolio Allocator agent to construct an optimal portfolio for each risk profile:

Crypto Portfolio Allocator · AI Agent · Nelly

Turn your crypto asset reports into a risk-aligned portfolio. AI agent by timothy.

It computes a risk-adjusted score for each asset from the reports, ranks them, and proposes a diversified allocation matched to the desired risk. The first portfolio took ~36 minutes (including parsing and scoring). The other two took ~5 minutes each since the parsing and scoring were already done. Here are the results:

Conservative

Portfolio Summary: Conservative

⚠️This portfolio summary was generated by the Nelly AI agent Crypto Portfolio Allocator without human review. It is for informational and educational purposes only and does not constitute financial, investment, or trading advice. Date: August 20, 2025 Investment Size: $1,000 Risk Appetite: Conservative Asset Selection LINK (25% - $250)

timothyTimothy Engqvist Johansson

Balanced

Portfolio Summary: Balanced

This portfolio represents a balanced allocation across 7 cryptocurrency assets, designed to provide diversified exposure to the digital asset ecosystem while managing downside risk through position sizing and sector constraints.

timothyTimothy Engqvist Johansson

Aggressive

Portfolio Summary: Aggressive

This aggressive cryptocurrency portfolio allocates across 12 carefully selected assets, emphasizing growth potential while maintaining diversification across the crypto ecosystem. The portfolio targets emerging opportunities in Layer 1 blockchains, DeFi protocols, and infrastructure assets.

timothyTimothy Engqvist Johansson

Costs

These are the LLM costs for allocating the portfolios:

• Conservative: $6.4 (includes report parsing and asset scoring)
• Balanced: $0.6
• Aggressive: $0.6

Reviewing the allocations

All the weights and amounts sum up correctly and the agent allocated the portfolios using the best-scoring assets. I reviewed the scores for the assets in the Conservative portfolio and the only possible error I found was that it included low liquidity / thin markets (+0.4) in the Risk Index for AVAX resulting in a lower Risk Adjusted Score (0.375 instead of 0.450).

Tracking performance

I'm tracking each portfolio against four benchmarks: BTC, ETH, and two crypto indices: NCI and COIN50. NCI is Nasdaq's Crypto Index, and COIN50 is Coinbase's index consisting of the top 50 leading crypto assets. The indices allow me to track whether my portfolios are performing better, worse, or the same as the crypto market overall. And comparing against BTC and ETH will tell me whether diversifying will have a better return than just holding BTC or ETH.

I'll update the performance weekly here:

Nelly Crypto Portfolio Performance

Google Docs

Next steps: Monitoring and Rebalancing

This isn't about day trading, so I don't want the agents making daily trades. The agent did recommend watching for stop-loss levels, major events, and other things that could justify rethinking positions. It also recommended monthly rebalancing, so I'm going to create a Rebalancing agent and perhaps a Monitoring agent to handle this.

Follow the progress

I'll keep posting updates on performance and on new agents I build for rebalancing and monitoring. Follow along here or on X to stay up to date. To the moon! 🚀