Workloft
▸ NOW LOADING WORKLOFT.LABS ◂

WORKLOFT LABS

Substrate before spectacle.

The research arm of Workloft. We track the AI frontier daily — 70+ papers a week, scored by Walt against nine substrate axes, citation-graphed via Semantic Scholar — and publish what to actually build for governed agent infrastructure.

▸ LABS API + MCP — get a free key All sections →
9846
Papers screened
5
Above threshold
9
Research axes
70
Notes published
▸ TODAY'S LAB BENCH
Five test tubes. Five papers. Top arXiv picks for today, scored against the Workloft research axes. Live from Walt's pipeline. Hover a tube. Click for the abstract.
▸ pulling today's picks from labs-api…
REG FIT = "would this clear an FCA Risk review, a UK GDPR DPIA, or a Local Authority procurement audit?"
●●● strong fit · ●●○ moderate · ●○○ low / academic-only.
▸ WORKLOFT LABS · OUTPUT · 11 PUBLISHED · NEW EVERY ~3 DAYS

WORKLOFT PAPERS

Substrate before spectacle.

Long-form research essays from the lab. One paper, one regulated lens, ~1,000 words each. Strong opinions, weakly held. Framed for FCA-regulated firms, UK Local Authorities and NHS Trusts — the buyers who have to defend the architecture, not the demo.

▸ THIS WEEK'S MUST-READ · PAPER №84

Orchestration Is the Easy Part of an Agent Fleet

Anthropic shipped Claude Code Projects: a coordinator, parallel cloud threads, shared memory. We have run that shape by hand for months. The coordinator was never the part that bit us. Shared memory staying true was.

~1,350 WORDS 7 MIN READ AGENT INFRA · ORCHESTRATION · MEMORY READ PAPER №84 →
▸ THE LAB

Pick a wing. Each one opens to its own page — methodology, ledger, problems, the pipe, the watch-list.

9×
The 9 Axes
The rubric Walt scores every paper on. Published openly because rubrics that don't survive sunlight aren't rubrics.
Methodology→
✓
Replication Ledger
Papers we've cited and actually rebuilt. What worked, what didn't, what the authors didn't write down.
Proof of work→
∂
Open Problems
Substrate-layer questions without published answers yet. Six on the bench. Drop us a line if you've hit one in production.
6 problems→
@
Who We Read
Ten researchers we follow personally. The list itself is a statement of taste before any output.
10 names→
→
The Pipe
Ingest → Score → Graph → Synthesise. How a paper becomes a Workloft signal.
Process→
⌘
Labs API + MCP
Programmatic access to the lab bench. HTTP endpoints + MCP server. Free tier live — get a key.
API · MCP · Free→
★
Substrate Score
Public 9-axis benchmark for AI agent runtimes. Vendors submit, Walt scores against the published rubric, the leaderboard is open.
Benchmark · v0→