How the way we work is shifting.

Less "new model dropped", more "the ground moved". Tracked over 7, 30 and 90 days.

🛠️ AI Development

Last 7 days — what's newly moving

Developers are codifying a repeatable operating model for coding agents instead of one-off prompting: one builds a workflow around a local task tracker so they can restart into fresh context at will; another defines each agent in a plain TOML file — prompt, model, skills, context files — piped in through git hooks and cron, no framework. The process, not the prompt, is the artifact.
The dev conversation flipped from can it code to what did it cost: a coding agent burns ~33k tokens per turn in fixed overhead before you type a word; one dev ran 50 tickets in 36 hours and 'burned my Max plan', another is 'running out of tokens, not ideas' after 41 subagents. A Gartner figure already puts 6% of firms over $2k/dev/month in tokens.
The response is turning cost into an engineering discipline: teams now track tokens per task like microservice calls, route dull fetch-and-format jobs to a cheap model instead of a reasoner ($10 gone in a morning before the fix), trim per-turn context, and pull work onto local models to cap the bill. One team cut agent token cost ~50% just by reading its own usage data.

Last 30 days — what's consolidating

Repo-resident context files have become the alignment mechanism: several engineers hand-write ADRs, specs or pattern notes first and let the agent execute against them — decision records in the repo instead of re-explaining the codebase every session.
Orchestration has not settled: some engineers describe collapsing multi-agent rigs into a single long-context thread with an upfront discussion phase, now that models finish whole tasks; others still run planner–worker–evaluator loops or parallel review sub-agents and prefer it. A live split, not a verdict.
💸 The token bills landed this month and split the room: one engineer calls a six-hour, 240-million-token, $125 run a bargain against months of expert work; on the other side sit a $700 unattended overnight loop with no stop condition, and an Anthropic bill jumping to $1.4m a year on a seat-tier cliff.

Last 3 months — what's held up

Agents as the default way to write code has held all quarter: the same shape keeps appearing since April — the agent implements, the human steers, reviews and merges — whether people run one tool for everything or a hand-built pipeline.
The cost pushback has been just as durable: a legacy-codebase experiment matched the reported $500–2,000 per engineer per month, and the bill does not stop at tokens — bandwidth, disk wear and battery drain get named too. The counter-current that never went away.
Reorganising teams around agents is real but contested: one widely-shared account of AI-native teams has engineers making product calls with far fewer PMs, while another reports companies that cut senior engineers and then paid contractors to rebuild the context that left with them. Our channel spent the month on the same question.
Keeping the exit open has hardened into practice: model routers and swappable harnesses so no vendor is load-bearing, context in local files the team owns — and people act on it, like long-time Cursor subscribers cancelling for cheaper Claude-over-SSH setups.

🎨 AI Design & UX

Last 7 days — what's newly moving

Packaging a visual language for agents to read is picking up speed: an open-source DESIGN.md of tokens that coding agents consume directly, a repo for teaching agents a unified design system, and the first paid AI-ready guideline packs. The same idea reached our channel twice this week.
The shape of the AI design tool is settling into something unglamorous: a folder of plain-text instructions that maps the workflow so the agent reads the right one at the right moment, plus a screenshot-review loop — the orchestrator captures the rendered page and hands it to a sub-agent to critique the UI, and each miss becomes a durable rule.
Solo operators are cutting designers out of commodity content: one founder describes a $500, one-day build of an AI that now renders every carousel and story their team ships; another swapped a $1,500 designer quote for a $20/mo tool plus minutes of editing. The craft end holds, though — a designer shipping 90% of his code via AI still can’t find a design tool he trusts for real product work.
As one-shot generation gets cheap, the value shifts to the critique loop: one designer argues taste alone yields mediocre output unless someone checks hierarchy and density before the agent races ahead — so references become explicit rules with rejection criteria set up front. Our own channel read the same week that taste can't be delegated, and picked up an open skill built to fight AI-slop UI.

Last 30 days — what's consolidating

Figma's job is being renegotiated, not ended: first drafts and exploration happen in AI tools, Figma keeps precision and final assembly. One designer wires their design system, component library and AI stack directly into it and iterates by screenshot feedback; another argues AI is exactly what makes it essential.
Design work keeps migrating into coding harnesses: a principal-level designer is described working 95% from a design.md in the terminal, and another builds one-off design tools by telling an agent what the tool should do. Our channel read the same shift — the T-shaped UX role giving way to a 'polymath architect'.
Demand is pointing both ways at once: one designer now wins clients because her files slot straight into their AI pipelines, while production teams are described going the other way — early optimism, inconsistent output, then quietly bringing designers back. Genuinely unresolved.

Last 3 months — what's held up

The personal multi-tool stack has held all quarter: separate tools for brainstorming, imagery, mockups and build, stitched together per designer — one triages fifty-plus generated mockups per screen by hand, another has AI study their own style before proposing layouts.
Skipping static mockups for working prototypes has stuck: a written plan goes straight to a clickable build, with AI used to think through flows, states and missed actions in the prototype itself rather than in frames.
The business arc of the quarter: an agency that trained design and PM agents on its own style now sells one-off system setups instead of retainers — while a persistent counter-current argues the token economics do not add up, with small-plan users describing quotas gone in three prompts.

🧭 Delivery & PM

Last 7 days — what's newly moving

Ticket boards are becoming agent interfaces: people describe hooking Linear up so the AI creates and manages the tasks itself, spawning sub-agents or assigning the manual ones back to the human — with one framing that PM tools now matter more for agents than for humans, because they hold the context agents need while humans move a layer up.
The sharper shift is agents that find their own work: one blueprint opens with a discovery loop that reads CI, issues and commits to decide what to do next; another sets an agent to scan CI failures and Jira automatically instead of waiting to be prompted. The PM's edge, one argues, now comes from delegating coordination early — the role moves from doing the work to routing it.
Even as boards turn into agent interfaces, a ceiling keeps showing: one engineer watched a capable agent stall on a decision not because it couldn't reason but because the context was scattered across Slack, a Jira ticket and Postgres with no way to reach it; another says hand an agent a real task and error recovery is the first thing that breaks. The optimists' answer is always-on autonomous queues.
A cost-and-ROI reckoning is reaching agentic delivery: one consultant watched a client kill a $1.4M document-automation project after 14 months — 12% time saved, error rate up — then hire process engineers who beat it by 45% for less. Others warn the spend goes non-linear across dozens of agents, and that long multi-step loops are exactly where per-step errors compound.
The ticket-to-PR loop is starting to close on its own for small teams: a bug in chat auto-files a Jira ticket, an agent on a Mac Mini implements it and opens a PR, a second agent reviews, and the founder just merges in the morning; elsewhere an AI 'teammate' scheduled its own cron and closed a ticket with zero human approval. Still hedged — even a builder running a dozen agents stops short of full dev handover.

Last 30 days — what's consolidating

Agent triage inside Jira is splitting the room: one account credits Atlassian's Rovo with clearing months of manual clicking, while others call it useless, describe two days debugging an agent that filed every ticket in the wrong project, and share it inventing a language-governance epic before fixing a typo.
For agent-driven work, Linear keeps beating Jira in first-hand accounts — lighter, cheaper, and with a more complete MCP for agents to drive.
One team ran 12,000 agent queries and found pre-joining Jira, GitHub and PagerDuty data into a single context lake came out 80% cheaper than wiring agents to each tool directly.

Last 3 months — what's held up

Pointing agents at the ticket backlog has held up all quarter: from telling an agent to work a Jira ticket — it pulls the task, plans, commits and moves it to review — to background loops grinding a Linear backlog, and a daily run that writes the morning stand-up from Linear, Sentry and git history.
Early but striking: at one AI-native company the first day of a new job is an agent workflow — dev environment set up, repos pulled, past Slack debates and Linear issues surfaced — with the new joiner describing feeling a year in after three days.

The day-to-day pulse The receipts — resources