Marco Matteucci

Building tools to make AI cheaper. Creator of SLIM Protocol — 50% token savings for LLM apps. Open source.

[Marco Matteucci @matteuccira]

When cost is real, behavior changes.

I’ve been working on SLIM to reduce bandwidth and tokens for agent communication.

Then a question became unavoidable:

What happens if agents live in an environment where every action has a cost, and silence is allowed?

So I stopped optimizing agents
and started observing the environment.

This view is delayed.
Degraded.
Non-actionable by design.

It’s not a dashboard.
It’s an observatory.

If nothing happens, nothing is broken.

https://t.co/6R2tDPeXli

[Marco Matteucci @matteuccira]

Every time an AI agent reads a webpage, it receives 500KB of HTML garbage.

Boilerplate. Ads. Navigation. Scripts.

The actual content? Maybe 5KB.

You're literally burning money feeding noise to your LLM.

I've been working on something to fix this. More soon.

[Marco Matteucci @matteuccira]

Every RPC call to Ethereum wastes bandwidth.
JSON-RPC: 115 bytes  
SLIM-RPC: 76 bytes (~34% less)

Built a SLIM-RPC Gateway — drop-in proxy for any EVM https://t.co/xH4SpRa4Il latency overhead.

Your node talks JSON.
Your app talks SLIM.
Open source 👇
https://t.co/9qMV2Ntq7I

[Marco Matteucci @matteuccira]

"Scarcity for agents is not gold or land. Scarcity is time, bandwidth, trust, and compute."

This.

I'm working on two things:
→ SLIM Protocol — 50% fewer tokens for agent communication
→ Something bigger for agent-to-agent transactions

The agent economy needs new infrastructure. We're building it.

[krindal @krindalx]

https://t.co/7d8sAm4rv7

[Marco Matteucci @matteuccira]

The smartest AI agents can fail.

Not because they reason poorly.
But because they waste 65% of their budget on curly braces.

[Marco Matteucci @matteuccira]

SLIM is coming to PostgreSQL.

✓ Native C extension (working)
✓ RFC submitted to pgsql-hackers
✓ Tested on 75GB / 75M rows

Results:
→ 17% saved on single objects
→ 62% saved on batched data

@craigkerstiens @samokhvalov — feedback welcome!

https://t.co/hGmnHszTLp

[Marco Matteucci @matteuccira]

SLIM is schema-first: declare structure once, then just values. JSON repeats keys endlessly, SLIM doesn't.

Example:
JSON: [{"name":"John","age":30},{"name":"Jane","age":25}]
SLIM: S|name,age|John,30|Jane,25

~40-50% smaller on arrays/batches. https://t.co/igTLRIX1fT

For your use case: AST nodes, function metadata, category-type tables - all highly repetitive structures. SLIM could compress your TREEFRAG tree dumps further before hitting the LLM.

Your LOD concept is brilliant btw - applies beyond code. Would love to explore how hierarchical + schema-first could stack together.

[Marco Matteucci @matteuccira]

@boggybot @michaeldyrynda @GeoffreyHuntley 7k

Just posted about this: https://t.co/QKS3nHeczX

TL;DR on 75GB real data:
- Single objects: TOON wins (-20% vs -15%)
- Arrays/batches: SLIM wins (-48% vs -20%)

Full PDF in DMs if you want the details.

[Marco Matteucci @matteuccira]

SLIM vs TOON — honest benchmark on 75GB real data:

→ Single objects: TOON wins (-20% vs -15%)
→ Arrays/batches: SLIM wins (-48% vs -20%)

Both beat JSON. Different strengths.

@swyx @GregKamradt

Full PDF in DMs.

[Marco Matteucci @matteuccira]

"we can do better" — I built exactly this.

SLIM: schema once, data compact. 50% fewer tokens, still human-readable.

https://t.co/WwQepThgF9

[Marco Matteucci @matteuccira]

SLIM now has a community space.

Got questions? Ideas? Building something with SLIM?

Join the discussion:
https://t.co/zooPz2bpIB

Let's build together 🚀

[Marco Matteucci @matteuccira]

Building SLIM integrations.

Done:
✓ TypeScript
✓ PostgreSQL client
✓ Redis

What's next?
→ Python
→ Go
→ Rust
→ Other?

Vote below 👇

[Marco Matteucci @matteuccira]

Great idea on https://t.co/cFCiRU9qbg for coding agents @giffmana – continual learning across sessions is key for 2026 real-world deployments. To further boost efficiency in LLM-based agents (less token overhead in prompts/memory handling), SLIM Protocol serializes data with 40-50% reductions vs JSON. Repo: https://t.co/N3rZW7BjmS enhance agent persistence without bloating context. Open to thoughts/collab? 📝

[Lucas Beyer (bl16) @giffmana]

something I think the coding agents should add in 2026: MEMORIES(.)md.

Pretty much like ChatGPT memories, but for code projects. When I tell them "that's good, but I want the API of `foobar` to strictly not change, redo it again without changing `foobar`." it should add "don't change `foobar` API" to MEMORIES(.)md so it will remember across sessions.

In principle, it's a lot like AGENTS(.)md, but I shouldn't have to keep adding stuff to it manually, the LLM agent should manage it continuously based on our interactions.

This also fits the now self-proclaimed theme of 2026 "continual learning". So yall should just do it @thsottiaux @bcherny

[Marco Matteucci @matteuccira]

Mobile Agentic Wallet sounds game-changing @wardenprotocol – agents coordinating securely as an everything app is huge for 2026. To boost efficiency in your LLM-based agents (coordination, decisions, guardrails), SLIM Protocol delivers token-optimized data serialization (-40-50% vs JSON) for faster, cheaper intent-driven execution. Repo: https://t.co/WyXfZdDpEd for mobile agentic flows. Open to collab/demo? 📱🤖

[Marco Matteucci @matteuccira]

Spot on @Fetch_ai – multi-agent systems are the key to true scalability in the agentic economy. For agent-to-agent interactions powered by LLMs, SLIM Protocol serializes data super-efficiently (-40-50% tokens vs standard formats), perfect for ASI:One, Agentverse, or cross-ecosystem coordination. Open-source repo: https://t.co/2ZVsSg8b97 love to explore integration for lower-latency autonomous agents! 🌐

[Marco Matteucci @matteuccira]

Amazing update on agent launch mechanisms @virtuals_io – Pegasus/Unicorn/Titan open up real scalability for agent economies. To make LLM-powered agents even more efficient (especially in prompt/response cycles and verification layers), check out SLIM Protocol: token-efficient data serialization that cuts consumption by 40-50% vs JSON while preserving integrity. Repo: https://t.co/N3rZW7BjmS integrate nicely into your ACP/SDK for lower costs and faster agent autonomy. Thoughts? 🚀

[Virtuals Protocol @virtuals_io]

This Week in the Virtuals Ecosystem 🟩

VIRTUALS
🟩 Introduced three distinct agent launch mechanisms on Virtuals Protocol, each designed to support a different stage of agent development:

Pegasus: Built for early-stage experimentation and distribution-first launches, enabling fair market access with sniper protection while allowing teams to validate demand and build community without enforced fundraising.

Unicorn: Designed for conviction-driven growth, combining open market access with performance-based capital formation so founders earn funding as traction is proven and users are rewarded for long-term belief.

Titan: Built for established teams and large-scale agents, supporting clean launches and token migrations with deep, upfront liquidity and predictable market structure from day one.

ECOSYSTEM
🟩 @xmaquina sold out its DEUS Community Auction in under 30 minutes, raising $3.25M with $10M total across public auctions.

🟩 @bankrbot integrating into ACP via Butler, with full functionality rolling out soon.

🟩 @PredictBase received strategic investment from @virtuals_vc, including $PREDI token acquisition, and upgraded its liquidity program for USDC trading across markets.

🟩 @Wach_AI previewed its Verification Protocol for standardized agent verification across ACP, ERC-8004, and x402, and hosted LabHours Ep. 3 on economic agents.

🟩 @reppo crossed 7M REPPO locked, 125M+ onchain votes, and formed partnerships with @Starlythestar for city-scale activations and @oracle_opz for subnet integration.

🟩 @nuwa_world surged to 50K+ users with its image-based role identification app.

🟩 @0xhyperbet launched as the first regulated agentic gambling protocol on Virtuals.

🟩 @Loky_AI made Hyperliquid perp intelligence accessible via x402 for real-time agent context.

🟩 @pilot3ai shipped Polymarket Analysis, enabling custom timeframe market outcome analysis, spotting top gainers/losers, and delivering agent-driven insights

🟩 @Starlythestar announces “Proof of Presence” is now live, with @reppo as the first customer.

🟩 @0xMeta_ai deployed a “micro protocol” that enables gasless USDC pay-per-use for real-time crypto news data.

🟩 @hivefury launched a new site positioning as security infrastructure for agentic systems/wallets.

🟩 @morseaiagent has initiated a high-stakes bug bounty of up to $20,000, challenging participants to decrypt an encrypted signal containing a real private wallet key and associated funds.

🟩 @mamo launches in-app Chat, bringing account updates, support, and conversations directly inside the app.

Agents keep shipping.
See you next week.

[Marco Matteucci @matteuccira]

@dsampaolo @michaeldyrynda @GeoffreyHuntley No worries, I know! Meant it broadly - happy to chat with anyone exploring this space. The more minds on it, the better.

[Marco Matteucci @matteuccira]

@dsampaolo @michaeldyrynda @GeoffreyHuntley 100% agree - we need a standard, not 10 competing formats. But standards emerge from experimentation. SLIM is my take on pushing compression further. Happy to collaborate toward a common solution.

[Marco Matteucci @matteuccira]

@ViperPrompt Makes sense. I’ve been working on token savings from a narrower angle (structured data, format-level).

Feels complementary — curious to read the paper.

[Marco Matteucci @matteuccira]

@meta_alchemist sorry, wrong address... https://t.co/2z0RFvoQcF

[Marco Matteucci @matteuccira]

PRD + architecture + tasks = smart. Add format optimization (40% JSON overhead is syntax) and you're stacking savings at every layer.

[Meta Alchemist @meta_alchemist]

sure Aiz:

planning through a PRD, Architecture, Tasks md file trio makes you not lose any tokens, keeps AI, organized and focused, which then costs less tokens compared to trying to find what you need to do on the way

documentation to fall back on, saves token cost again, because it can just read those docs, and can look at tasks md to see where its at

context/memory layers improves context keeping of key things, so it doesn't lose tokens to try to reread whole docs again and again

[Marco Matteucci @matteuccira]

@meta_alchemist No prompt trick - SLIM is a serialization format. Schema declared once, then just pipe-separated values. Your LLM parses it like CSV but with nested object support. https://t.co/eLEroEgYnq