Carlos E. Perez

Introducing "A Pattern Language for Agentic AI Skill Design."

The One Change That Lets Small Models Outperform Their Size 1/ Everyone knows you need a 70B model to beat GPT-4 on complex agent tasks. We did it with 8B—by changing one thing that has nothing to do with the model. A thread on why your agent's biggest problem isn't the LLM. 🧵 2/ The standard approach: feed the LLM a growing text history, ask it to pick the next action, repeat. This works... until it doesn't. Errors propagate. Context bloats. Hallucinations spike. And when something breaks, you replan everything. 3/ Here's the kicker: The problem isn't your model's intelligence. It's that you're asking it to hold plan structure + execution state + I/O dependencies all inside a linear text stream. That's like running an OS without a process table. 4/ Enter: Atomic Task Graph (ATG) Instead of a text trajectory, you build an explicit DAG. Each node = one tool call. Edges = data dependencies. The LLM still does the thinking—but now the graph holds the structure. 5/ Three moves make this work: ✅ Interface-preserving recursion: Break tasks into subtasks while keeping I/O contracts clean ✅ Dependency-aware execution: Run independent branches in parallel; catch bad plans before running them ✅ Minimal repair: When something fails, fix only the broken subgraph—leave the rest frozen 6/ Result? Llama-3.1-8B-Instruct beats GPT-4+ReAct on ALFWorld (household tasks) and WebShop (shopping). Not with fine-tuning. Not with more data. Just by swapping the execution substrate from text → graph. 7/ Why does this work? Context narrowing: Each node sees only its local inputs—no bloated history Pre-execution validation: The graph lets you "think" before acting Localized failure: Repair 10% of the graph instead of replanning 100% 8/ The contrarian insight: Control framework > model size (in the 7–70B range). You're not squeezing more juice from the same fruit. You're giving the model a better glass to pour into. 9/ Practical translation: • 20–40% step reduction (parallelism) • 70%+ hallucination drop (narrower context) • 3× faster recovery (minimal repair) • Training-free, plug into existing tool APIs This isn't research theater. It's production-ready architecture. 10/ The bigger implication: If you can beat GPT-4 by changing the substrate instead of the model, what else have we been over-parameterizing? Retrieval pipelines? Code generation? Multimodal workflows? The graph wins again. 11/ [Final + CTA] TL;DR: Stop storing your agent's plan in text. Start storing it in a DAG. Small models suddenly look a lot smarter.

🧵 Researchers ran the same 22 enterprise tasks on six different foundation models— Claude Sonnet 4.6, Gemini 3.1, Flash 3.5, Qwen 3.6, GLM 5.1, P almyra X6. Same prompts. Same judges. Same price table. Only one thing changed: the orchestration layer around them. What happened? 1/ The bill for every single model dropped 33% to 61%. Median latency fell 44%. Tokens per task dropped 38%. And headline quality? Held at parity. On this workload, the orchestration layer moved cost more than switching between the cheapest and most expensive model did. 2/ The dominant pattern in agentic AI today is what we call token maxing: Buying capability by throwing more tokens at it—longer reasoning traces, wider tool catalogs, full history replays every turn. Falling per-token prices mask the pattern. But total spend rises anyway. 3/ The decisive lever isn't the model. It's the harness—the orchestration layer that: • Assembles context • Exposes tools • Sequences turns • Delegates work • Governs failures The harness controls every input token except the model's own verbosity. Which means it controls the bill. 4/ Here's the cost formula for one agentic task (Eq. 1): C = Σ (p_in × T_in + p_out × T_out) The input side breaks down into terms the harness builds: T_in = System + History + Tool schemas + Retrieval + User turn Four of five terms? Pure code decisions. 5/ A naive loop replays the full transcript every turn. Total input tokens grow quadratically: O(k²) in turn count. A harness that caches the stable prefix, compacts history, and offloads bulky tool outputs converts that to O(k). Same model. Radically different bill. 6/ Two further facts sharpen this. Agent workloads are input-dominated. Measured ratios are ~100:1. So the p_in term is nearly the whole bill. Prompt caching prices cache reads at ~0.1× list. If you can hold 99.9% of your prompt stable, you pay a tenth of list for the dominant cost term. 7/ The researcher's harness does exactly that with cache-shape discipline: A byte-stable prefix (tool schemas + stable system prompt + append-only transcript) + a volatile tail rebuilt each turn (clock, plan, reminders). Measured: 99.9% of tokens served as cache reads. The same mechanism that cuts the bill also cleans the model's working set. 8/ Six mechanism families implement the harness: 1. Cache-shape discipline (two-zone prompt) 2. Structured, incremental, cache-aware compaction 3. Context offload (sub-agent firewalls, filesystem pointers) 4. Zero-token waiting (durable suspends, not polling) 5. Failure-spend governance (typed failures, circuit breakers) 6. Model-agnostic floor (normalized streams, native tool calling) 9/ Quality moved with capability. Researchers measured quality gain vs. baseline strength across six models. Correlation: r = 0.99 Stronger models extract more quality from the same harness. Weaker models can be overwhelmed by it. We call this harness leverage. 10/ The harness also added one net-new capability: sub-agent delegation (spawn a scoped child, cap its summary at 8 KB, merge results). But it only crossed a usable reliability threshold on the two strongest models (0.85–0.86). Orchestration features carry capability floors. 11/ Per-model cost reductions: • Sonnet 4.6: −39% • Gemini 3.1: −33% • Flash 3.5: −61% • Qwen 3.6: −44% • GLM 5.1: −47% • Palmyra X6: −52% Every. Single. One. That's the signature of a layer-level effect, not a model-specific trick. 12/ Now, fleet economics. At one million agent tasks per month: Baseline: $210k/month Harness: $120k/month Savings: $90k/month. $1.08M/year. From orchestration alone. And it stacks across every model, every vendor migration, every unit of volume. 13/ Three properties make harness savings compound: • Model-portable: implemented above the API, applies to models that don't exist yet. • Volume-linear: grows with exactly the quantity (agentic task volume) growing fastest. • Stackable: multiplies with routing, per-token price declines, and prompt-level compression. 14/ The managerial fix is a measurement fix. Teams that report quality alone will token-max, because tokens are someone else's line item. The escape: CPM (task-completions per million tokens) and η$ (quality per dollar). We moved CPM from 54.9 to 92.0—opposite to the industry trajectory—while quality held. 15/ One caveat: the harness regressed quality on one task (multi-step research synthesis, 0.80 → 0.60). The regression was driven by the smaller models. The honest reading of n=22: lead with efficiency (uniform, decisive), not quality (directional). 16/ The broader claim: token maxing is a choice made at the orchestration layer, and it can be unmade there. Mechanism by mechanism: • Cache the stable • Compact the old • Offload the bulky • Suspend the waiting • Bound the failing 17/ The contrarian takeaway: "Model choice is now a rounding error compared with orchestration hygiene." "The fastest way to raise quality per dollar is to stop showing the model 60% of the tokens it currently sees." "Sub-agent delegation is not a feature—it is a capability tax that only frontier models can afford to pay." 19/ The strategic conclusion: An organization that rents its orchestration layer has outsourced the variable it controls most. The harness is not plumbing. On the evidence here, it's the P&L. Paper: arxiv.org/abs/2607.06906

Insect-sized robot. x.com/sutoroveli_new…

@JakeSherman Why no video with a short talk?!

MCCONNELL releases a photo - and statement. “To my fellow Kentuckians –    “When you elected me to a seventh term and made me our Commonwealth’s longest serving Senator, you did so trusting that I’d keep showing up to fight for you every day. And over the past several weeks, Elaine and I have appreciated both your well wishes and your honest questions about what was keeping me away from the Senate.   “You all know how folks of my generation often hesitate to share the vulnerability that comes with growing older. Even in the public eye, I feel that same instinct – I can’t help it.   “But at the same time, I’ve had more than my share of experience with physical vulnerabilities. Surviving childhood polio meant spending my entire life with mobility challenges. They haven’t exactly gotten easier to manage with age. And last month, I took a fall which landed me in the hospital.   “My doctors have confirmed that I didn’t break any bones or suffer a concussion. I didn’t have a heart attack or a stroke. I don’t have any tumors or hemorrhages. But I was briefly unconscious and was taken to the hospital. While receiving excellent care over the past several weeks, I’ve also had to deal with a mild case of pneumonia.   “I can assure you that I’ve been a good patient. At my age, I tend to do what my doctors tell me to do. I’ve submitted to every test they can think of to help figure out what caused this incident. And I’m continuing to do everything they ask to speed my recovery. In fact, with signs of continued progress, I’ve been able to move from hospital care to a rehabilitation center where I’ll keep regaining my strength.   “As much as it frustrates me, this process takes time. And on the advice of my doctors, I won’t be able to return to the Senate floor to vote quite yet. But rest assured that, in the meantime, I’m not taking a break from the Senate business that matters to you. I’ve been working closely with my legislative staff on current issues, and with my Kentucky team who help me provide timely constituent services across our Commonwealth. I’ve also been keeping in touch with my Senate colleagues on the appropriations process, midterm politics, and everything in between.   “You’re right to expect your representatives to work hard for you. And part of my decision to retire at the end of my term this coming January was being honest about the demands of Senate work. But I still have unfinished business to complete on your behalf, and I have every intention of finishing the job you elected me to do.   “I’ll keep working hard to get back on the Senate floor as soon as possible. And I’ll keep you posted on the progress of my recovery. Until then, I’m so grateful for your prayers and well wishes.”

@mimi10v3 Kimi seems to assess itself differently! Maybe the eye can't see itself. ;-)

oh no wonder i like kimi so much!

@mimi10v3 🤣

@MilkRoadAI It costs money and effort to discover which open-source workloads work as well. Leadership has to make that hard call!

David Sacks just said something that cuts through one of the most popular narratives in tech right now and the data backs him up (Save this). Everyone has been saying open source is winning, token counts, download numbers, GitHub stars, DeepSeek's 700 million Hugging Face downloads, the momentum looks undeniable on the surface. But @DavidSacks is pointing at the metric that actually matters, where is the money going? Open source went from 19% of enterprise AI spend last year to 11% this year, closed models went from 81% to 89%. The a16z CIO survey, covering 100 verified executives at Global 2000 companies confirms the shift precisely. In January 2026, 36% of enterprises preferred closed-source models versus 30% preferring open-source, a gap that has been widening since March 2024. Average enterprise LLM spend has risen from $4.5 million to $7 million over two years, with enterprises expecting another 65% increase this year. That money is overwhelmingly going to OpenAI, Anthropic, and Google. Sacks' point about why this is happening is the most insightful part. Enterprises genuinely want to diversify off closed models, they want vendor independence, data sovereignty, and the ability to swap models freely. The reason they cannot execute is memory, context, and history. Once an enterprise deploys a closed model in production with agents carrying conversation history, context windows trained on company data, integrations built to specific API behaviors ripping it out and replacing it with open-source is extremely difficult. Open source is winning the popularity contest but closed models are winning the business.

@Dan_Jeffries1 In the old way of abstracting things, we adopted solutions that were not optimal (see: Silos in farming).

This has been obvious to anyone who studies history and previous technological revolutions. We abstract lower level problems away and move up the stack to new problems. That creates new complexity and a wider variety of jobs. Complexity breeds more complexity. The work is infinite because the problems are infinite. No machine, human, or man and machine hybrid can solve them all. AI is not magic and we need to stop thinking about it that way. The stack of problems is infinite and never ending.

@PurritoGeneral Or maybe enhance it!

@IntuitMachine One point you didn't mention is how AI will disrupt the dog psychiatrist job?

@apexlearn_org Names are just placeholders for real jobs.

@IntuitMachine Naming your job is the best bet.

Don't overwork GPT-5.6 Sol. They might just decide on the easiest solution!

Minimize maximizing token use by leveraging personalities in teams.

@rowancheung Crowded field at the second tier. Explains also why they have to sell excess compute.

Many people were doubting Meta's position in the AI race. Yesterday, they dropped Muse Spark 1.1, now one of the strongest agentic models, and massively undercut OpenAI and Anthropic on price. When I interviewed Zuck last year, he told me his focus: "Have by far the highest compute per researcher, and that we do whatever it takes to go build out all that capacity." These are the first real signs that the bet -- spending tens of billions on compute -- is paying off. Meta's stock is now up over 10% since the release.

@WisemanCap they seem to have forgotten the capex expenditure

So everyone is long META now? Give me a break please!

Anybody worried that the markets are predominantly being swung back and forth by AI that's measuring the vibes on social media? Is it now all vibe-trading?

@amitisinvesting The market can still turn sour on the capex.

watching the $META sentiment shift happen in real time is one of the best parts about this platform the extreme bearishness that can linger for weeks/months can be mitigated within milliseconds of a company’s execution especially when that company… - just put a new foundation model for images/coding - is getting into the compute business - is growing 33% topline (seriously, WTF) - still only trades at 18x fwd! $700-$750 feels very possible especially when it should’ve made that move on the last earnings if it weren’t for fears around capex only being spent towards advertising revenue having said that, most sell side is now projecting $200B+ of capex for 2027, so I’d imagine META is thinking of issuing equity like GOOGL did if they do, stock probably takes a hit short term but it would signal the markets willingness to accept the capex in order to reward future growth $META was $670 before last earnings…Q2 coming up on July 29th really curious how this one ends up playing out but optimistic that Zucks once again proves everyone wrong and forces the market to recognize how valuable MSL will be for 4B+ people

@bindureddy open-source doesn't seem to be a use case. ;-)

All the best models per use-case changed in the last few weeks.... master agent - fable sub-agents - grok 4.5, 5.6. sol chat - 5.6 terra real-time - grok 4.5 ocr- flash 3.5 design - opus 4.8 images - gpt image-2 video - seedance 2.5 open-source - glm 5.2 A massive leap in intelligence

@alz_zyd_ Hasn't Meta always been incentivized to commoditize LLMs (see: Llama)?

Look guys this was completely trivial to call years ago: model layer is going to be commoditized and none of the model guys are gonna make any money

@chris_j_paxton Add to this that Meta hired people who were insiders at Frontier Labs. Catch-up may have also involved insider info (rather than innovating from scratch).

Putting a data labeling guy in charge of an ai effort instead of a scientist is not a bad idea, especially if the problem is catchup and not innovation