← /editions
#72

The harness is the expensive part

// TrueFoundry open-sourced TrueForge with a claim worth testing: hold the model constant, change the harness, and identical agent work costs 30% less. Alongside a local-first agent and a portable skill pack, the pattern is clear — the model layer is commoditizing and the wrapper is where cost and lock-in live.

$ today_in_10_seconds

TrueFoundry open-sourced its agent harness yesterday with a claim worth reading twice: hold the model constant, change the wrapper, and the same work costs 30% less. Swap in an open model and it's 75%. Two other releases today — a fully local agent and a portable SEO skill pack — point the same direction. The harness, the layer nobody benchmarks, is where cost and lock-in actually live.

$ lead_story

TrueFoundry open-sources TrueForge, and the numbers are about the wrapper

TrueFoundry — a San Francisco infra startup founded in 2021 by ex-Meta engineers — released TrueForge on August 19 under the MIT license on GitHub and PyPI. The benchmark is 14 cross-system tasks from DevRev's Enterprise-Bench, spanning a CRM, an issue tracker and a document store: with Opus 4.8 in both harnesses, each completed roughly 11 tasks, but TrueForge averaged $8.50 per run against $11.80 — about 30% less. Point TrueForge at GLM-5.2 instead and the average drops to $2.90 for comparable completions, roughly 75% off. Out of the box it ships 40+ built-in tools, sandboxed execution, human-in-the-loop approvals and automatic context compaction, with model and MCP traffic routed through gateways that enforce rate limits, budget caps and RBAC. Read the benchmark with the appropriate squint: TrueFoundry ran it themselves, and the results haven't been independently validated at production scale. But the shape of the claim is the interesting part, not the decimal places. If a third of your agent spend is determined by how the harness compacts context, batches tool calls and retries failures, then your model bill has been measuring the wrong thing. That's testable on your own workloads this week.

// why_it_matters: If the same model costs 30% more depending on who wraps it, "which model" was never the expensive decision.

read more →

$ also_worth_attention

  • Magnitude — an agent that brings its own inference engine

    Magnitude is an Apache 2.0 agent that runs local models with nothing else to configure: it profiles your hardware, recommends and downloads models, then runs them inside the agent — no Ollama, no model server, no inference setup. macOS and Linux native, Windows via WSL, installed with a single npm global. The pitch from founder Tom Greenwald is blunt: today's agents are local but the model isn't, so every prompt, file and secret ends up at Anthropic or OpenAI. The quiet detail matters more than the privacy pitch — it loads standard skills, including Anthropic's own xlsx, pptx, docx and pdf packages, straight from the skills registry. Same skill files, different runtime, no model vendor involved.

    link →
  • SEO Skills AI — a domain skill pack that installs into eight harnesses

    An MIT-licensed suite of 27 modular sub-skills and 18 specialist sub-agents that install into Claude Code, Cursor, Windsurf, Copilot, ChatGPT/Codex, Cline and Antigravity from one script, with a deterministic zero-telemetry Python execution engine and first-party adapters for DataForSEO, Firecrawl, Bing Webmaster and IndexNow. The security posture is more thought-through than most: user-supplied URLs pass through a fail-closed check that blocks private and metadata IPs and re-checks redirects, the shipped Cline config doesn't auto-approve fetch tools, and the remediation engine prints patches for review rather than applying them. Temper the enthusiasm, though — the repo is at 13 stars and 11 commits, and its own "empirical results" screenshots are captioned as illustrative UI examples rather than packaged proof. Treat it as a well-built template, not a validated system.

    link →
  • SpaceX reportedly went after Cognition — and Cognition says it didn't happen

    Bloomberg reported Wednesday that SpaceX approached Cognition AI about an acquisition, that Cognition didn't engage, and that talks continue about a compute arrangement instead. Cognition CEO Scott Wu disputed the report on X within hours, saying the company isn't for sale and that the two haven't been in talks. The approach would have come days after SpaceX's $60B Cursor acquisition closed, and Cognition — whose Devin agent serves Mercedes-Benz, Citi and Goldman Sachs — is in early talks to raise at $40B or more. Wu's denial didn't address the compute-partnership claim. Note what's being priced here: not models, but agent products with enterprise customers attached.

    link →

$ bigger_picture

The trend behind the headlines

For two years the model was the product and the harness was glue. Today it inverts. TrueForge separates the runtime from the model and finds a third of the cost hiding there. Magnitude drags the model down onto your laptop and keeps the skill format intact. SEO Skills ships domain expertise as portable files that work in eight harnesses nobody has to agree on. And SpaceX is reportedly bidding tens of billions for companies that sell harnesses, not weights. The pattern: the model layer is commoditizing into something you swap per task, while the layer around it — how tools are called, context is compacted, permissions are enforced, cost is capped — is where the differentiation and the switching cost accumulate. If you're still choosing a vendor by benchmark scores, you're negotiating over the cheap part.

$ tool_of_the_day

skills.sh

The package manager for agent skills. `npx skills find` to discover, `npx skills add` to install; skills land in `.claude/skills/` or the equivalent for whichever agent you're running. It has been used to install skills across amp, antigravity, claude-code, codex, cursor, droid, gemini, github-copilot, goose, kilo, opencode and roo. If today's thesis holds and your harness is now a swappable component, this is how you keep your accumulated instructions from being locked to any one of them.

explore skills.sh →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime