Four agents share one repo, and every write must score 0.82 to land
Four Medusa agents move chips across a shared board, where each write gets a CRC check and is discarded below a 0.82 score; the author calls it a lower-level version of the Jev engineering pattern.
holy sh*t. 404.1K lines written, 214/214 evals, canary at 100%, and only 3 of 8 guards lit 4 agents write into one repo here, so every write has to pass a score before it touches the board if score < 0.82: board.discard(block) 4 Medusa agents seek, grab and carry chips across one shared board. 20 chips are placed by the end every chip read gets a crc check, and a mismatch raises an error instead of landing writes go through lanes on a 64-bit bus. the board scores each block and keeps only what clears 0.82 The board only needs a human at three critical moments: → before a chip lands: did the crc match, or is an agent writing from a corrupted read? → when a block scores under 0.82: is discard() dropping noise, or throwing away the one fix that mattered? → before prod: the deploy bar says 100%, so who signed off with 5 guards still dark? 4 agents share one board, and the caption puts prod last This is the Jev engineering pattern I keep coming back to, one layer down. In 22 seconds lines written go from 182.2K to 404.1K and shipped goes from 911 to 959, while the deploy log cycles through trace, a2a handshake, eval, canary and guard The crc check and the 0.82 score sit on every write. Deploy is the last panel on the board. - the full board setup > 4 agents: Medusa 01 to 04, each on its own seek, grab, carry loop > read_chip(): await http://chip.read(bus=64), raise IOError on crc mismatch > weave(): async for block in stream, score it, inject or discard under 0.82 > build order: jev-core, chip-loader, sig-router, code-weaver, eval-gate, deploy-ring > prod gate: evals, guards, deploy %, cost, with p95 latency and error rate on screen paste this prompt into Claude Code below: "Set up my Claude Code project so parallel subagents can't write unchecked code into the same repo: 1. Give each subagent its own worktree or branch, never the shared main checkout 2. Add a merge gate to CLAUDE.md: > every subagent change runs tests and lint before it is merged back > a change that fails is discarded and reported, not patched over 3. Add a PreToolUse hook on Bash that blocks deploy and push unless the last full check run passed 4. Order the build: core first, loaders and routers next, eval gate before any deploy step 5. Before asking me to approve, show: > which agent wrote what, the failing checks, and the final diff Show every configuration diff first. Do not apply edits until confirmed" the clip is a render of this pattern, not a recording of a live run
This post quotes another post: x.com/mr_buzzoni/status/2103689373774483815