Igor Negrutsa

Exploring practical AI systems, agents and tooling in large production environments.

Crypto enthusiast.

[Igor Negrutsa @Irns]

Tested the new dynamic workflows in Claude Code on a couple of real project tasks today.

The most interesting part was not the autonomy itself, but the quality improvement coming from orchestration and parallel review flows.

One case:

before a large refactor I needed stronger test coverage around a very large legacy component.

Most of the implementation work had already been done using a more traditional multi-agent/review workflow.

I had:
- a detailed plan
- existing implementation changes
- additional tests
- separate review flows

The new dynamic workflow feature was then used as a higher-level verification/orchestration layer to review the overall direction, coverage quality, and architectural consistency.

Instead of simply generating "more tests", the workflow launched multiple parallel agents that analyzed:
- existing tests
- dead code
- uncovered critical paths
- refactor intent
- actual business-risk areas

One particularly interesting part was that the workflow was not simply spawning "more agents".

Different agents were reviewing the task from different verification perspectives:
- migration relevance
- live vs dead code paths
- coupling risks
- regression gaps
- architectural intent
- test value vs test noise

That review quality felt significantly better compared to a single long-running agent session.

Another case involved reviewing and refining a custom agent workflow + skills setup for my project.

Again, multiple agents reviewed the architecture, decomposition logic, and implementation details from different angles before the execution flow continued.

A lot of issues and weak spots were caught surprisingly well.

Both runs were expensive (~1M tokens each), but the quality difference was very noticeable.

Feels like this kind of orchestration starts becoming much more interesting once:
- tasks become large
- context becomes noisy
- architecture matters
- verification/review quality matters more than raw generation speed

This feels much closer to workflow engineering than classic prompting.

https://t.co/bIj3vhZRkH

[Igor Negrutsa @Irns]

Still trying to fully understand and explore the idea of “Harness Engineering”, but my current impression is that the industry focus is slowly shifting:

from prompts → to context → to execution environments around agents.

Very simplified:

- Prompt Engineering = how you ask the model
- Context Engineering = what information the model receives
- Harness Engineering = how the entire execution workflow around the model operates

The interesting part is that once agents become more autonomous, the main problem is often no longer the prompt itself.

It becomes:
- task decomposition
- execution loops
- delegation
- retries/recovery
- tool orchestration
- context lifecycle management

One thing that feels increasingly real in practice:

more context does not always make the model smarter.

Long-running overloaded sessions often become noisy, inconsistent and less reliable over time.

This seems to be one reason why many agent workflows are moving toward:
- smaller bounded tasks
- isolated execution flows
- cleaner context windows
- structured orchestration around the model

Not “one giant smart agent”, but systems of smaller focused workflows working together.

Still learning this space, but the shift feels very real.

[Igor Negrutsa @Irns]

Feels like this approach could probably be adapted for other agent tools too.
For example Cursor or Codex with custom request/response parsers and tracing layers.

[Igor Negrutsa @Irns]

Wanted to better understand what actually happens under the hood in Claude Code sessions, so I tried cc-wiretap.

Really useful tool if you want to inspect:
- the real system prompt
- tool definitions
- full messages[] payloads
- raw API flow
- token/cache usage
- what actually gets sent to Anthropic

What turned out even more interesting was not the debugging itself, but the visibility it gives into why the agent behaves the way it does.

For example:
- why context suddenly explodes
- which instructions/files actually become part of the final context
- how linked context and references get expanded
- how much hidden context/tooling actually accumulates
- what the model really sees before generating a response

One particularly interesting thing was being able to inspect built-in commands like /compact and see the actual prompts/workflows behind them.

That kind of visibility makes it much easier to:
- understand how agents, tools and skills interact together
- analyze reasoning and context flow
- experiment with memory systems
- optimize prompts/tooling layers
- reuse ideas for your own commands and workflows

--debug and cc-wiretap are actually different layers of visibility.

--debug is great for:
- reasoning loops
- hooks
- MCP debugging
- tool execution flow

cc-wiretap is better for:
- raw API payloads
- system prompts
- tool schemas
- full request/response visibility

Using both together gives a surprisingly complete picture of how Claude Code actually operates internally.

Feels like this kind of tooling becomes extremely valuable once you start building larger agent workflows.

[Big Brain AI @realBigBrainAI]

Lisa Su (CEO of AMD) unveils the world's smallest AI development PC, capable of running 200B parameter models locally.

[Igor Negrutsa @Irns]

@LucaNetz Las Vegas Sphere?

[Igor Negrutsa @Irns]

@pudgypenguins GM

[Igor Negrutsa @Irns]

@degennQuant Has anyone asked him about 500mln $ tokens unlocks per month starting from November 29?

[beeple @beeple]

WARREN'S RETIREMENT PLAN

[Igor Negrutsa @Irns]

@FartCoinOfSOL Price )

[Igor Negrutsa @Irns]

@cz_binance $SNAI, $BAD

[Igor Negrutsa @Irns]

@badmeme_coin @SimonsCatMeme Where and how claim SBT?

[Haseeb ＞|＜ @hosseeb]

My 2025 Crypto Predictions

I'm either going to look like a prophet or an idiot over these predictions, but one thing is for sure: I'm going to piss off a lot of people with bags.

Breaking this up into six sections: my predictions for L1s/L2s, token launches, stablecoins, regulation, "AI Agents" (oh boy), and crypto x AI.

~9 minute read!

1. L1s/L2s
- The distinction between L1s and L2s is collapsing. Users no longer perceive the differences between L1s and L2s (did they ever?). The blockchain landscape, L1s and L2s combined, is overcrowded and due for a shakeout. The consolidation will be less about technical superiority—it will be about having a unique niche and building stickiness through GTM.

- Despite the strength of SVM and Move, EVM market share will actually grow in 2025. This growth will be driven by @base, @monad_xyz, and @berachain. This will be not because of compatibility anymore—it'll be because EVM/Solidity just has way more training data, and LLMs will be writing most of the application code in 2025. Already having a deep library of battle-tested cryptography contracts will also be a separator, because LLMs suck at writing low-level code. DevEx and footguns will matter less than training data and solid libraries in the LLM era of development.

- Solana will pressure more blockchains to optimize for low latency. We will move from TPS wars to latency wars—infra like @doublezero and super low-latency L2s like @megaeth will push user expectations toward web2 responsiveness. Expect more embrace of optimistic UIs, preconfirmations, intents, email onboarding, in-browser wallets, and progressive security. Shoutout to @privy for advancing the meta here.

- @HyperliquidX has demonstrated that specialized chains can work when they're laser-focused on a specific application and prioritize UX and easy bridging. More projects will follow this model. The old dream of one chain to rule them all is dead.

2. Token Launches
- The current meta of everyone doing huge airdrops via points programs is over. We are moving to a two-track world.

- Track one: if a project has a clear north star metric, like an exchange or a lending protocol, they will distribute tokens purely off points. They will not care if they are farmed or gamed—they are effectively distributing the token as a rebate/discount on the core KPI of the protocol, and the farmers are your actual users anyway.

- Track two: projects without clear north star metrics (like L1s and L2s) will move toward crowdsales. They may do smaller airdrops to reward social contributions, but the majority of tokens will get distributed via crowdfunding. Airdropping for vanity metrics is dead. Those aren't really going to users, they're going to industrialized farmers.

- Memecoins will continue to lose market share to “AI agent” coins. I consider this a migration from financial nihilism to financial over-optimism. (Yep I'm coining that.)

3. Stablecoins
- Stablecoin usage will explode, particularly among SMBs. Not just trading and speculation—real businesses will start using on-chain dollars for instant settlement.

- Banks are noticing: expect to see announcements of bank-issued stablecoins toward the end of 2025. They will not want to be left behind. But especially with Lutnick as Secretary of Commerce, Tether will remain #1.

- Expect @ethena to gobble up even more capital, especially as treasury yields continue to decline over the coming year. When the opportunity cost of capital declines, it makes basis trade yields even more attractive.

4. Regulation
- Stablecoin legislation passes in the US, while the broader market infrastructure overhaul (FIT21) gets delayed. Stablecoin adoption accelerates while Wall Street adoption, asset tokenization, and other TradFi integration will lag behind.

- Under Trump, Fortune 100 companies will become more willing to offer crypto to consumers, with tech companies and startups showing higher risk appetite. Trump's inauguration will create a perceived regulatory jubilee until clear rules and enforcement priorities are set. During this window, expect to see aggressive expansion of crypto integration into Web2 platforms.

5. AI Agents (this is the longest section because my thoughts here are likely controversial—read to the end!)
- The “AI agent” craze will continue probably throughout 2025. But it will die off eventually. This is not the long-term disruption to watch out for from AI, but it will be CT's fixation because it is the most social.

- These things are not really agents. These are chatbots with memecoins attached; they are barely agentic at all besides posting on Twitter. Current "AI agents" are also mostly "Wizard of Oz" agents—there are humans behind the scenes ensuring the AI doesn't go off the rails. This won't change any time soon because current agents are too janky (even Fortune 100 companies are not using agents in prod yet). Current agents can easily be manipulated into saying crazy things that damage their brands, or can be jailbroken to steal all of their resources. See @freysa_ai for what an actual autonomous AI looks like—if your favorite AI is not getting jailbroken, it’s because it's a Wizard of Oz AI.

- That said, I think this trend will accelerate. Chatbots can indeed replace a lot of influencers because chatbots never sleep, they're always on-message, and they’re less greedy than human influencers. Plus the majority of influencers aren’t very original anyway. Real-time information aggregation/amplification can be easily replaced by an algorithm even today (see @aixbt_agent).

- Right now these chatbots are fascinating to us because they are so novel. It’s like seeing an elephant paint. The first time you see it, you don’t really care that the painting is not very good—it’s spectacular to see. But the 1000th time, the novelty wears thin. I believe that will start to happen as these chatbots plateau.

- You can see that today with aixbt—it’s already pretty good at aggregating data about different projects. By next year and the next generation of agents, maybe aixbt will hallucinate a little less, go a little deeper, have a little smarter takes. But how much will you even notice? It’ll probably feel the same to most people.

- I think this novelty and market eagerness continues throughout 2025. Crypto takes a while to get bored of the shiny thing. But by 2026 I think there will be a sudden reversal. The chatbots will become so ubiquitous that people will get turned off by them. Sentiment will reverse. Seeing stories of their favorite human KOLs losing their livelihoods will kindle a kind of class consciousness. Users will start discriminating in favor of human KOLs, even if their content is less consistent.

- In response to this pro-human bias, chatbots will start hiding that they are AIs, trying to pass as humans in order to capture more of the attention market. Instead of monetizing through memecoins like today, future chatbots will monetize the same way human KOLs do—through sponsorships, affiliate links, and pumping tokens they own. KOLs will be routinely accused of being chatbots, and you will see AI-unmasking scandals. This will all get weird.

- But there's a darker side yet. Remember, LLMs are currently great wordcels, but not great at the other stuff yet. What are the best ways to make money as a wordcel in crypto? First is being an influencer, sure, but close second is being a scammer. You will start seeing autonomous scambots proliferate. These will explode, comparable to what ransomware and cryptojacking became post 2017. Expect this to become a real social problem.

- But while chatbots are likely to remain the center of attention in 2025, the long-term disruption from AI will not be at the social layer.

- And no, it’s not going to be in trading either. AIs will not give everyone their own “trading agent” or miniature hedge fund. Yes, AIs will scale everyone, but they will scale people proportionally to their capital, data, and infrastructure. You should therefore expect AI to supercharge preexisting trading firms who have capital scale and data scale. In other words, trading firms will become even better at making all of the money. It will also collapse the hierarchy among trading firms (most of them will become comparably good, since everyone will have access to 150 IQ quants in the cloud).

- Over time, AIs will make markets extremely efficient—even smaller, niche markets—which will leave little edge left for normal traders, even with their little homebrew assistant AIs. The value of original research will plummet. That said, the increased competition and liquidity should be a boon to the rest of us who are injecting noise into the market. (It will also mean @Polymarket liquidity on everything!)

- So if the big story is not chatbots and not trading bots, what else is there? Here’s my core thesis, which for some reason almost nobody is talking about: the truly impactful AI agents will be software engineering agents.

- Why is this such a big deal? Ask yourself this: what is the primary input to our industry? What is the costly input preventing there from being more applications, more wallets, better infrastructure, better everything? The answer is software. If AI agents cause the price of software to collapse, that will change everything.

- In a post-AI era, instead of having to raise millions of dollars for a seed round, you will be able to launch an application with $10K of AI cloud compute. Self-financed projects like Hyperliquid and Jupiter will go from the exception to the norm. The amount of applications and experimentation on-chain will absolutely explode. For an industry that is driven by software, this deflationary shock is going to lead to an on-chain renaissance.

- The implications of this on security are profound. AI-powered static analysis and monitoring will become ubiquitous, making security more accessible to everyone. These AIs will be fine-tuned on EVM/Solidity or Rust codebases, trained on vast databases of security audits and attack vectors. They'll be RL’d in simulated adversarial blockchain environments. I’m increasingly convinced that AI tools ultimately favor defenders over attackers when it comes to security. You will have AIs constantly red-teaming contracts, while other AIs will be hardening them, formally verifying their properties, and honing their skills at incident response and remediation.

- In the meantime, sure, trade AI-flavored memecoins. But real agents are going to have a lot more impact than tweeting and pumping their own tokens.

6. Actual Crypto x AI
- Above I detailed the impact of AI on crypto (which is the primary direction of influence), but crypto will also have an impact on AI.

- Truly autonomous agents will use crypto to pay each other. This will be especially true once there are permissive stablecoin regulations—you’ll start seeing even large companies that run AI agents using stablecoins for agent-to-agent payments because they’ll be so much easier to spin up than bank accounts.

- We will also see more and bigger scale experimentation around decentralized training and inference. A new generation of promising projects like @exolabs, @NousResearch, and @PrimeIntellect will pave the way for real alternatives to centralized training and company-owned models. @NEARProtocol is also going all-in on trying to create a full-stack credibly neutral and permissionless AI stack.

- The other place where crypto and AI will intersect is UX. Post-AI wallets will be completely transformed—an AI powered wallet should be able to take care of bridging, optimize trade routes, minimize fees for you, paper over interoperability issues or frontend bugs, and steer you clear of obvious scams or rugpulls. You won’t be juggling between multiple different wallets and changing RPCs or rebalancing your stablecoins—the AI will handle all of it for you. This likely takes until 2026 to become reliable enough to transform crypto’s UX. But when this arrives, what does this do to blockchain network effects? What happens when users stop caring—or even experiencing—which chain an application lives on?

- This space is still young, but I’m hopeful we’ll see things take off here soon. In the long run (say mid 2026) I expect this will be where most of the market cap of “AI x crypto” lives.

---

That’s all I got for predictions. I promised I'd write this before I hit 100K followers, so I'm a little late, but still within the new year!

Happy New Year everyone. Looking forward to being out of a job by this time next year! 🫡

Disclosure: These are all my personal opinions and do not represent the opinions of Dragonfly; Dragonfly holds investments in many of the names I mentioned in this piece. Not financial advice. DYOR. I may or may not be an AI.

[Igor Negrutsa @Irns]

@pudgypenguins Sorry, but I faced with bug, I linked 3 wallets 2eth and 1 sol but claimed was only tokens for sol wallet 50%.
etw wallets tokens marked as claimed but it's not true.
Tried to clean up site data but didn't help.

[Igor Negrutsa @Irns]

@RealFlokiInu What happened on MEXC? I spend 360$ for Monkey, and now it's 8$ 😤🥺

[Igor Negrutsa @Irns]

@CetusProtocol @hippo_cto I can't find HIPPO-SUI pool

[Igor Negrutsa @Irns]

@CryptoHayes $CAT

[Igor Negrutsa @Irns]

@CCommunityAI Экосистема скорее мертва чем жива , капа большая , на 3х продал и забыл

[Igor Negrutsa @Irns]

@CCommunityAI А какая польза от токена кроме голосования?

[Igor Negrutsa @Irns]

@CCommunityAI @symm_io твиттер скор просто огромный