Editorial
The noise around AI agents and robot demos is loud, but the day’s clearest signals come from engineering: how teams run critical systems, how defaults shape developer trust, and which open projects are plumbing the next wave of infrastructure. Today’s top pick is a rare look inside production engineering at scale — and the rest of the digest follows the same theme: durable systems beat flashy launches.
Top Signal
A couple million lines of Haskell: Production engineering at Mercury
Why this matters now: Mercury’s postmortem on running "a couple million lines of Haskell" shows how strong typing, workflow durability and instrumentation scale in real fintech production — useful for teams rethinking reliability under regulatory and growth pressure.
Mercury’s engineering write-up argues that the payoff from a heavy‑use functional language isn’t purity for its own sake, but operational clarity: types and disciplined interfaces become tools operators use to reason about running systems, not just compile-time niceties. They’ve leaned on replayable workflows (using Temporal and an open hs-temporal-sdk) to replace brittle cron/state-machine glue, and they instrument libraries so every business action is observable and controllable.
“Purity is a boundary, not a property,” the post summarizes, and the team treats the type system as “an operational aid” to preserve institutional knowledge.
Those design choices matter because fintech operates under stress—regulatory scrutiny, bursts of transaction volume, and the need to reproduce complex flows during incidents. Mercury’s practical tradeoffs are concrete: encode invariants in types where silent corruption is costly, but keep pragmatic escapes (documented uses of unsafePerformIO) when necessary. They also emphasize shipping observability alongside APIs so on-call teams can actually understand and stop runaway work.
If your team is building durable backend systems, the immediate takeaways are specific: prefer replayable, idempotent workflows; standardize observability inside libraries; and use type-level boundaries to push complex, imperative operations behind narrow, well-tested interfaces. For organizations scaling fast, those patterns buy you fewer firefighting nights and clearer post-incident narratives.
Source: the Mercury engineering post at A couple million lines of Haskell.
---
AI & Agents
No stories in the AI & Agents beat cleared our quality threshold today for deep treatment. The space remains noisy with demos (humanoid robots, consumer rollouts) and many early agent tools, but the lasting signals are still about governance, safe defaults and auditable behaviour — not shiny product videos. We’ll keep watching for agent‑safety work that pairs capability with provable controls.
Markets
Markets had a lot of chatter but few high‑quality, lasting beats today that met our threshold. Broad themes persisted — AI infrastructure flows into chip and memory demand, and energy/geopolitics keep making markets asymmetric — but the durable engineering and governance signals came from developer and ops stories below.
Dev & Open Source
VS Code inserting 'Co-Authored-by Copilot' into commits regardless of usage
Why this matters now: The VS Code git extension change default flipped behavior that can automatically add AI co-author metadata to commits, creating consent, attribution and auditability headaches for millions of developers.
A schema default change (git.addAICoAuthor flipped from "off" to "all") made a small config tweak have outsized effects: hosts or test environments that don’t load contributed defaults can still pick up the new behavior, producing surprising commit trailers like:
"Co‑authored‑by: Copilot"
That matters beyond noisy commit history. Commit metadata feeds audits, provenance tools, and even licensing workflows; a default that annotates authorship without explicit consent or clear opt-in breaks expectations and can complicate traceability in supply‑chain investigations. The incident is a reminder that defaults are policy: when tooling authors expose behavior that affects legal or audit signals, the bar for silent defaults needs to be higher.
Source: the original PR and discussion at VS Code PR #310226.
Do_not_track: a single env var to opt out of telemetry
Why this matters now: The DO_NOT_TRACK env‑var proposal offers a simple convention — “export DO_NOT_TRACK=1” — that could immediately reduce surprising telemetry from CLI tools, if maintainers adopt it.
The write‑up frames the env var as a command‑line analogue to NO_COLOR: one lightweight signal users set once to stop non‑essential telemetry, telemetry‑based ads, and remote pings. It’s pragmatic and implementable: authors get a single check in boot code; users get a universal opt‑out.
“export DO_NOT_TRACK=1”
Adoption is the catch: commercial incentives will resist a universal opt‑out, and an env var can itself become identifying metadata. Still, for sysadmins, privacy‑conscious devs and CI builders, the idea is low friction and worth pushing into major SDKs — if enough projects honor it, it reduces surprising network calls during builds and test runs.
Source: the proposal at do_not_track.sh.
dav2d — an open, fast AV2 decoder in progress
Why this matters now: VideoLAN’s dav2d aims to provide a fast, portable AV2 software decoder to fill the hardware gap, which can materially speed AV2 adoption across browsers and players while silicon catches up.
dav2d is presented as a spiritual successor to dav1d for AV1: a BSD‑licensed, cross‑platform C decoder focused on speed and correctness. The repo is early and the authors warn “should not be used in production,” but its roadmap (C decoder, then hand‑tuned asm for AVX2/ARM) mirrors the path that helped AV1 reach wide playback earlier.
For product teams shipping video, this matters because a robust software decoder reduces reliance on hardware offload and accelerates compatibility across devices. For contributors, the project needs C/assembly work and testing help; for vendors, it’s a sign that open software can unlock codec adoption faster than waiting for silicon.
Source: VideoLAN’s dav2d repository at code.videolan.org/videolan/dav2d.
Ladybird browser’s April engineering rollup
Why this matters now: Ladybird’s monthly update shows how thoughtful plumbing work — parsing, threading, allocator choices — makes a hobby engine feel usable on real sites, lowering the bar for browser diversity.
The team shipped hundreds of PRs: inline PDF rendering, improved address‑bar suggestions, bookmark management, and under‑the‑hood gains like speculative HTML parsing and off‑thread bytecode compilation. These aren’t flashy features; they’re the technical fixes that reduce site breakage and improve speed, and they matter because alternative engines face a compatibility moat more than a feature gap.
Source: Ladybird’s April update at Ladybird – April 2026.
---
The Bottom Line
Today’s signal is operational: the teams that win under stress are those who instrument, standardize defaults carefully, and treat code as infrastructure. Mercury’s Haskell playbook, the VS Code metadata surprise, and proposals like DO_NOT_TRACK all point to the same lesson: discipline in defaults, observability, and workflow durability is the competitive advantage engineers can actually control.