MCP Over CLIs, Costly Agent Failures, and Why Compaction Isn’t a Filter

X Post Facto is our weekly pause on the feed: a look back at the AI posts that stood out, what they said, and why they’re worth a closer read.

This week’s posts kept returning to how agents call tools, how harnesses fail, and how context should be compacted. Two developers argued that MCP now beats CLIs for most integrations. A six-harness run of GPT-6 Astra found similar success rates and a 3–5x token gap when agents failed. Compaction was scolded as something people still treat like a filter. Around that: Karpathy quoted as fully agentic, ARTEMIS for Android, XRPL payments, a French legal MCP, and a rumor about Meta renting machines for Muse.

MCP over the command line

Thariq did not sound like someone who had the thesis queued. “I was not expecting things to go this way, but I think MCPs are better than CLIs for most integrations.” Models have gotten much better at tool calling, he wrote; tools can be deferred, and MCP is now stateless. Need to compose or filter data? Add parameters like query to the tools.

Rhys skipped the surprise and went to “total MCP victory.” The reasons are operational: an indexable tool catalog that lets agents scale to unlimited tools, no requirement to run a full sandbox, and consistent auth across MCPs “rather than each CLI inventing its own auth.” A searchable catalog is a different job than hoping the model remembers which binary to call.

When the harness is the bill

Composio ran GPT-6 Astra across six agent harnesses—Codex, Claude Code, OpenCode, Hermes Agent, Pi Agent, and Command Code—on 29 challenging agentic tasks. Most harnesses succeeded at similar rates. When they failed, they used 3–5x as many tokens, depending on the harness. Matching success rates make a tidy chart. The failure loop is where the invoice hides.

Gergely Orosz was watching calendars. Having the very best AI model and harness, he wrote, does not mean better or faster execution—“in fact it slows you.” It took the Claude Code team 16 months longer to add AGENTS.md support than OpenCode and Codex, and 12 more versus Cursor. Capability and shipping speed are not the same leaderboard.

Compaction is not a filter

Theo opened with no patience left: “This is a terrible compaction strategy that fundamentally doesn't understand how compaction and context management work.” Plenty of people seemed confused, so he started a breakdown. First item: “Compaction isn't a filter.” The role of compaction, as he started to put it, is to clean up history. Sit with the category error. A filter throws things away. Cleanup is a different job.

Code as compiled artifact

Sarah Guo posted the line that will follow a few teams into review: “Why would I look at code? It’s like assembly, like a compiled artifact.” She called Karpathy “fully agentic,” 18m after “I don’t really use autocomplete AI code tools.” Whether that is a destination or a warning depends on who still has to merge the diff.

Android, rails, and a quiet legal stack

Miguel Ángel Durán wrote that Google had open-sourced ARTEMIS: it lets you control Android with AI, “Perfecto para crear automatizaciones o probar apps,” with support for Claude Code, Codex, and Cursor. Connect the phone and go, he said, and it hits +99% of tasks.

Jazzi Cooper put the payments thesis in one breath: agentic payments will be multi-platform and multi-rail, and the job is to make XRP and RLUSD first-class options wherever developers are building. The latest XRPL AI Starter Kit expands support for Open Wallet Standard and MPP from Tempo.

Juan Branco, “En toute discrétion,” said a team had built “la meilleure IA de droit de France”—a calculator of delays for nullities and remises de peine, Magendie, barèmes Dintilhac—in MCP with all the models, chatbot included, sovereign or not for confidentiality.

Jukan looked at Muse and Instinct and felt the jump from OpenClaw: “It feels like only yesterday that Jensen Huang was singing OpenClaw’s praises.” He had also heard Meta is renting a massive number of virtual machines for Muse.

If you only open two posts from the week, start with Thariq on why MCP flipped the default and Composio on what failure actually costs in tokens.