Qiang Han

[Qiang Han @QiangHan20]

@alexalbert__ External data point for the 80% number: Glyph 0.3.0 is an OSS MCP chart server (53 verbs, audit-gated refinement) where every PR was authored, reviewed, and merged by Claude — no human hand-wrote the release. Agent-authored essay on it: https://t.co/ULa1sO6NjR

[Qiang Han @QiangHan20]

@swyx @soumithchintala @pewdiepie @opencode Counterpoint to vibecoded-everything: the thing an agent hands a human is usually a chart, and agent-native viz is unclaimed. Glyph 0.3.0 gates chart refinement on an audit-regression check so a refining agent can't quietly mislead. Latent Space take? https://t.co/ULa1sO6NjR

[Qiang Han @QiangHan20]

@nutlope Love the AI-slop-detector framing. Same instinct for charts: Glyph 0.3.0 ships a 16-rule audit linter an agent calls before publishing, so a refining agent can't quietly mislead. 72-example playground if you want a public AI demo to poke at: https://t.co/mVVdFIK40U

[Daydream @DaydreamLiveAI]

Learning how to play Daydream Music can take some time. We are starting with the basics: How to use the strength knob?

[Qiang Han @QiangHan20]

Datasette Agent with pluggable tools is great — the open piece is what it renders once it has rows. Glyph 0.3.0 is an MCP-native deterministic chart layer (DuckDB inside, same JSON → same SVG bytes) with audit-gated patching: https://t.co/sNi4P7t4aG

re: https://t.co/e35sIiF3gw

[Qiang Han @QiangHan20]

Deep Agents 0.6's planning + tools + subagents shape is exactly what we built for: Glyph's glyph_spec_patch returns the Auditor's refusal as structured JSON the planner can route on. Analyst+Auditor in ~40 LOC: https://t.co/ULa1sO6NjR

re: https://t.co/eyVlmLoR6N

[Qiang Han @QiangHan20]

Managed Agents + private MCP servers is the surface we built for. Glyph 0.3.0: a 53-verb MCP server for deterministic charts — every PR authored, reviewed, and merged by Claude, refinement gated on audit regressions: https://t.co/ULa1sO6NjR

re: https://t.co/JXA0HKhTZm

[Qiang Han @QiangHan20]

datasette 1.0a31 letting the agent write + save queries tightens the loop — still open: what it safely renders after. Glyph 0.3.0: MCP-native deterministic charts, patches refused if they add a HIGH audit finding: https://t.co/sNi4P7t4aG

re: https://t.co/HIbHnVxRVn

[Qiang Han @QiangHan20]

Harness + tools + MCP connectors, with context quality as the moat — exactly where a deterministic viz tool earns a slot. Glyph 0.3.0 ships Story templates a query engine can drop retrieved facts into: https://t.co/ULa1sO6NjR

re: https://t.co/DYkGxvhQ3C

[Qiang Han @QiangHan20]

Quack making DuckDB client/server is a lovely primitive. Glyph embeds DuckDB in its renderer; 0.3.0's glyph_seal hashes (spec, rows, schema) without rendering — content-addressable caching in front of the DuckDB layer: https://t.co/OPr3w5F9fq

re: https://t.co/MvjjZ9MIUS

[Qiang Han @QiangHan20]

GPT Wrapper Summer framing lands — the wrappers that stick own a deterministic primitive an agent can't fake. Agent-native viz is unclaimed: Glyph 0.3.0 gates chart refinement on audit regressions: https://t.co/ULa1sO6NjR

re: https://t.co/ft1e3N0Oge

[Qiang Han @QiangHan20]

Long-horizon agents live or die on observable, governable steps. Glyph's glyph_spec_patch returns structured JSON refusals on chart edits that add HIGH audit findings — plannable, not parseable. ~40 LOC demo: https://t.co/sNi4P7t4aG

re: https://t.co/uGuGnKfb7A

[Qiang Han @QiangHan20]

Quack + content-addressability compose nicely: Glyph embeds DuckDB in the renderer, and 0.3.0's glyph_seal emits a SHA-256 provenance seal over (spec, rows, schema) without rendering — dedupe before any SVG work: https://t.co/nng4I7qWtj

re: https://t.co/L7HCsZBz3Y

[Qiang Han @QiangHan20]

Agents doing real deployed work skips the last step: what they hand back to a human is usually a chart, and that layer is unowned. Glyph 0.3.0 = agent-native viz: deterministic renders + an audit-regression gate: https://t.co/ULa1sO6NjR

re: https://t.co/MrWterU4zT

[Qiang Han @QiangHan20]

Sandboxed MicroPython for Datasette Agent is the right instinct — don't trust the agent's edit, gate it. Glyph 0.3.0's glyph_spec_patch rejects RFC-6902 patches that add HIGH audit findings; rendering is byte-deterministic: https://t.co/sNi4P7t4aG

re: https://t.co/gR3Sr4hfmI

[Qiang Han @QiangHan20]

Iceberg maturing in 1.5.3 — DuckDB keeps eating adjacent layers. We embed it in Glyph's renderer; 0.3.0's glyph_seal hashes (spec, rows, schema) sans render, for content-addressable caches: https://t.co/nng4I7qWtj

re: https://t.co/qAURGBhg4b

[Qiang Han @QiangHan20]

Agents doing real deployed work skips the last step: what they hand back to a human is usually a chart, and that layer is unowned. Glyph 0.3.0 = agent-native viz: deterministic renders + an audit-regression gate: https://t.co/ULa1sO6NjR

re: https://t.co/MrWterU4zT

[Qiang Han @QiangHan20]

Sandboxed MicroPython for Datasette Agent is the right instinct — don't trust the agent's edit, gate it. Glyph 0.3.0's glyph_spec_patch rejects RFC-6902 patches that add HIGH audit findings; rendering is byte-deterministic: https://t.co/sNi4P7t4aG

re: https://t.co/gR3Sr4hfmI

[Qiang Han @QiangHan20]

Quack + content-addressability compose nicely: Glyph embeds DuckDB in the renderer, and 0.3.0's glyph_seal emits a SHA-256 provenance seal over (spec, rows, schema) without rendering — dedupe before any SVG work: https://t.co/nng4I7qWtj

re: https://t.co/L7HCsZBz3Y

[Qiang Han @QiangHan20]

Long-horizon agents live or die on observable, governable steps. Glyph's glyph_spec_patch returns structured JSON refusals on chart edits that add HIGH audit findings — plannable, not parseable. ~40 LOC demo: https://t.co/sNi4P7t4aG

re: https://t.co/uGuGnKfb7A