📈 Trends

Radar · Trends — the trajectory. How ways of working with AI are shifting in development, design & UX and delivery & PM, distilled from what people say about their own work on X (every post verified) and what our community shares. Counter-takes kept in.

Updated 14 Sep 2026 · X posts found via Grok and verified against the originals · community shares from our own archive.

🛠️ AI Development

Last 7 days — what's newly moving

updated 14 Sep 2026

• The engineering stays human — the agent just executes it. A recurring stance this week: one 30-year engineer uses assistants to implement his engineering, not be the engineer; another assigns GitHub issues straight to a coding agent in a workflow he still owns; a third keeps durable rules in an AGENTS.md, apart from each feature’s throwaway context. The gripe is letting the AI do it for you.@eas4ai ↗@prasadhonrao ↗@georgehanu ↗
• Less agent — now a cost argument. Because the agent loop re-reads its whole context every turn, the bill tracks context size more than model price, so the trims turn concrete: 36.5% off a benchmark from a two-line cut to skills, one harness taken 4GB → 600MB, bloated MCP tool-lists pruned. Counter: one dev tallied $13.5k for 58 days of agents — you cannot cap a ceiling you never measured.@nicholasjconn ↗@dheerajbatra ↗@btsouth ↗@memselon ↗@bytecrafter_1 ↗counter:@rodgui ↗@ventry089 ↗📚 CLAUDE.md Is RAM, Not Disk📚 Use ASD-STE100 to Stop LLMs Writing Slop Documentation
• Running a fleet of agents is becoming its own discipline. People share setups for driving multiple agents in parallel, and agents now hand work to other agents — one notes Qwen Code documents Codex or Claude delegating tasks to it (you → agent → agent → subagents). The much-shared caveat: a multi-agent setup is a coordinated distributed system whose single point of failure has to eat lunch.@undefinedKi ↗@gadi_neelesh ↗counter:@cto_junior ↗

Last 30 days — what's consolidating

updated 14 Sep 2026

• *“Context engineering” is now the named discipline — and split over where context should live. The frame: most agent failures aren’t prompt failures, so the work is choosing what the agent sees at each step. One camp hand-writes specs, rules and config into the repo; another argues the reverse — context lives outside the repo, and the job became what can it reach* (plugins into GitHub, Slack), not phrasing.@YoussefHosni951 ↗@DmitroCP ↗@MystiqueMide ↗@heinapurola ↗📚 CLAUDE.md Is RAM, Not Disk📚 OpenSpec: Spec-Driven Planning for Coding Agents📚 Multi-Agent Coding Without Worktree Chaos
• 'Loop engineering' is settling in as the name for the discipline. The frame: stop babysitting — assign, wait, review, fix — and build the loop itself, owning the parts you control. As one put it, an agent is four parts in a loop; you own one, and the other three are what break it — so the work moves to designing the scaffolding, not prompting turn by turn.@0xPure_eth ↗@hanakoxbt ↗📚 Loop Library📚 Loop Engineering📚 The Art of Loop Engineering
• 💸 The cost conversation is hardening from shock into policy: some firms now cap engineers at a fixed weekly token budget (~$100), and the bill's hidden second half — reasoning tokens, retries — is described landing months after go-live. The emerging fix is architectural: pull the repetitive, deterministic steps out of the agent and only call the model for the genuinely dynamic part.@svpino ↗@v_shakthi ↗@DGlushakov41949 ↗@arizeai ↗counter:@Kavyabuildss ↗
• How far to trust agent code is becoming a per-file call. The most-shared take: decide the tradeoff per file, not a blanket policy — you can’t both ship fast on AI code and grasp every line. Output is real: one team measured 51,300 lines of code plus 43,900 of tests. Counter: engineering is far from solved — one dev, three days in, still hit obvious bugs; another calls agent work more all-consuming.@arpit_bhayani ↗@IngoElfering ↗counter:@mehulmpt ↗@AVatchinsky ↗

Last 3 months — what's held up

updated 2 Jul 2026

• 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.@TheShortBear ↗@housecor ↗@jessethanley ↗📚 My Agentic Coding Workflow
• 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.@ChemistDeFi ↗@janstevens ↗@marty_kausas ↗
• 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.@AndrewYNg ↗counter:@bbxjasper ↗📚 AI should shrink teams📚 Is AI making teams better?
• 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.@GergelyOrosz ↗@mclzion ↗@migmartri ↗@dustinhyle ↗

🎨 AI Design & UX

Last 7 days — what's newly moving

updated 14 Sep 2026

• The design system is becoming something an agent _runs on_, not a Figma library. This week’s workflows share a spine: nail the system in Claude, then have a harness reference DESIGN.md, canonical examples and versioned tokens each run — one syncs canonical colours to Figma to stop drift, another turns finished screens back into DESIGN.md. The aim: not a prettier draft but the same output every time.@abrahamabel196 ↗@matt_teeixeira ↗@TimothyReavis ↗@_studable_ ↗📚 awesome-claude-design: 68 DESIGN.md files to stop AI apps looking identical📚 You Don't Have a Design System📚 Building an LLM-safe design system (Polar's Orbit)📚 Hallmark: An Open Design Skill That Fights AI-Slop UI
• The design handoff is shifting from a flat file to a clickable build. A couple of product designers describe dropping static Figma mockups for coded prototypes — one ships Cursor + shadcn flows so stakeholders click through and UX gaps show up earlier; another builds dozens of screen iterations with an AI coding tool, opening Figma late if at all. The bet: a prototype people can use beats a picture.@RoyQuilor ↗@kvncnls ↗
• The design job is tilting toward orchestration. The AI-first front half is reframed as owning product context and a plan — capturing customer calls and requirements, running ideas through Opus or GPT, then drafting as if handing to a dev team, the traditional tool opened last for hi-fi. If designers are orchestrators, what are we orchestrating? one asks. Counter: one stays manual, AI only to wireframe.@dianne_alter ↗@shedsgns ↗@LordMarket22 ↗@ThomasBurleson ↗@vadym_petryshyn ↗@Anwuriii ↗counter:@ivabuzuk ↗📚 Your First 30 Days With AI in Your Design Workflow
• *AI is moving to the front of the process — testing directions before a sprint. Designers use it to test ideas earlier and refine with better context, not polish too soon; a prompt or screenshot now spins up multi-screen editable wireframes to check a flow before anyone commits. Counter: it still one-shots mediocrity — one keeps iterating with their own eye*, since AI learned the internet’s average.@bestofyasi ↗@nyike ↗counter:@deeplakhani ↗

Last 30 days — what's consolidating

updated 14 Sep 2026

• Figma's role is being renegotiated inside a live loop with Claude, not ended: designers describe generating screens, flows and tweaks in the AI, which writes them back into the file and builds the prototype — 'me, Claude, and Figma in a loop.' The counter-current is sharper this month — some have stopped opening Figma entirely since Claude Design shipped.@DannPetty ↗@maybepratikk ↗@benlaudesign ↗@phena_designs ↗counter:@sxtvik ↗
• A new job for the design system: being readable by agents, not just humans. The complaint recurs — 'your agent has no idea what your design system looks like, so every project ships as AI slop, Inter showing up unasked.' Teams are responding by encoding tokens, rules and structure so agents build on-brand — and report the architecture calls getting harder even as production gets faster.@chris_bgp ↗@gavinelliott ↗@pixelmatters_ ↗📚 Building an LLM-safe design system (Polar's Orbit)📚 Design System Architecture: Theming, Experiments, and Keeping It Maintainable📚 Hallmark: An Open Design Skill That Fights AI-Slop UI📚 UX-Context Design: Using UX Knowledge to Steer AI-Generated Design
• Design and engineering are collapsing into one role. Work keeps migrating into the terminal — a principal-level designer is described working 95% from a design.md — and design teams now say the two disciplines 'stopped being separate.' The archive tracks the same arc: the T-shaped UX specialist giving way to a polymath architect, and a constellation of new hybrid design roles.@pixelmatters_ ↗@petergyang ↗📚 T-shaped UX → polymath architect📚 A Constellation of Emerging Design Roles📚 The Terminal Became My Canvas
• *A new seam in visual-design work: generate, then edit the elements, not re-roll. A couple of designers describe the same pipeline — make a concept in an image model, then have an agent hand it to Photoshop and split it into layers*, so one element changes without regenerating the whole image. Early, only a handful of accounts, but a concrete fix for generative design’s gripe: no control over details.@viipin8 ↗@TheoBuildsAI ↗
• A single but loud counter (~2.7k reactions): one product designer’s “I hate building frontends with AI.” The tools generate a dashboard in 30 seconds — then it’s six hours explaining that the main button is invisible, the table breaks on mobile, and every screen is the same pile of rounded cards. The last-mile craft gap hasn’t moved, even as generation speeds up.counter:@aryanXmahajan ↗

Last 3 months — what's held up

updated 2 Jul 2026

• 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.@thattechangel ↗@KayaHickin ↗@creativemints ↗@yancacchione ↗
• 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.@jp ↗@nickbakeddesign ↗@newincreative ↗
• 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.@abnux ↗counter:@edzitron ↗@atomtanstudio ↗

🧭 Delivery & PM

Last 7 days — what's newly moving

updated 14 Sep 2026

• The delivery lead’s coordination rituals are going agent-run: a scrum agent captures blockers and status for async standups, grooms the backlog and trims retros; another pipes Slack and meeting notes into RAID items, decisions and first-draft exec updates. The job moves to vetting that output — and the span widens, one describing 5–10 projects at once where it used to be one or two.@mugabuilds ↗@shirshakchavan ↗@basilboy ↗@rileybrown ↗@YallaSaikalyan ↗📚 Inside PM at Stripe: What Comes After 'Become a Builder'
• As leads point more agents at the work, prioritising across them becomes the new job: one runs a formal priority-queue framework — levels, aging, dashboards, overrides — so low-value work stops blocking critical work; another maps a dispatch matrix to specialist agents. The counter is first-hand: more agents just add noise, give one the lead — and AI only runs a broken process faster.@AiCamila_ ↗@alex_prompter ↗@sonxpiaz ↗counter:@taelo_kim ↗@ContinuityPm ↗
• As the busywork automates, the bottleneck moves up — not away. The role is being redrawn: less chasing status and building reports; more setting priorities, clearing blockers, aligning people. But the honest catch: when creation time collapses (a strategy paper in 20 minutes), decision time rises — someone still reviews every call — and AI cannot change the operating model underneath.@shirshakchavan ↗@priyavr ↗counter:@barryoreilly ↗📚 Inside PM at Stripe: What Comes After 'Become a Builder'

Last 30 days — what's consolidating

updated 14 Sep 2026

• The ROI reckoning has arrived in delivery: a 16-month automation build scrapped as useless after a few hundred K, a founder '$2,400/month on AI tools and the team's still drowning,' and the month's most-shared line — AI just recreated the oldest org chart: one expensive planner, cheap labour doing the work. The emerging answer isn't more tools but process-first — fix the workflow, then automate.@PJMorell ↗@Sergio__bm ↗@mengxi ↗@lukepierceops ↗@businessbarista ↗📚 Is AI making teams better?📚 AI should shrink teams
• The delivery function itself is reorganising around agents. Accounts describe post-sales turning into delivery — 'the software only creates value when the work gets delivered' — and the lead's job becoming orchestration: pointing a PM-agent at a team of coding agents and letting it route the work. Early, mostly from AI-native shops, but consistent.@andreslo ↗@draginol ↗@edgarpavlovsky ↗@rileybrown ↗@AlexFinn ↗@sharepointfocus ↗📚 Inside PM at Stripe: What Comes After 'Become a Builder'📚 The Human-in-the-Loop Is Tired
• AI sped up creation, so the bottleneck moved to review and coordination — still human work. A much-shared account: AI decoupled effort from output, so a dev ships a giant PR in minutes that a lead needs hours to vet. Others say the jams are organisational — one change needing nine teams to align — and agencies now build in pushback after clients skip discovery and lose real money (one lost $20k).counter:@housecor ↗@v1becoder ↗@michal_gren ↗

Last 3 months — what's held up

updated 2 Jul 2026

• 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.@VicVijayakumar ↗@wildsatchmo ↗@shaun5k_ ↗
• 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.@jmwind ↗