Activity runs execute on Temporal in every environment — durable, resumable, restart-safe — with run history that tells the truth.
Reliability week:
- Temporal everywhere — activity runs execute on the same durable engine in local, dev and production. A run survives a deploy mid-flight and resumes where it stopped.
- Honest run history — every run row states what triggered it, what happened and what it produced. One run is one card; re-runs re-scan instead of silently reusing stale data.
- Human-in-the-loop, fixed — the agent asks once and only for values it genuinely needs, renders choices as buttons, and never re-surfaces a question you already answered. An empty answer blocks the run rather than being guessed past.
- Graceful deploys — in-flight runs are given time to drain on redeploy instead of being killed.
- Engineering: test suite ~10× faster — the full backend suite (1,700 tests) now runs in about two and a half minutes; database schema is provisioned once per worker and cloned per test.
Score rings on every audit card, Templates page with instant sample previews, and reports that generate themselves after a run.
The reports library got a face and a memory:
- Report cards — every audit card shows its headline score ring and key metrics at a glance, straight from the run's real data.
- Auto-generated reports — when an audit activity finishes, its paired report is generated automatically; no extra click.
- Refresh & download — re-render any report from the latest data (without re-running the whole workflow) and download it from the pane header.
- Templates page — see every deliverable design and brand master in one place; each HTML design previews instantly with sample data, even in a brand-new workspace.
- Open in chat — every report opens into a refine chat, so iterating on a deliverable is one click.
Logos, palettes, fonts and brand voice pulled from a client's site — validated before they're stored, honest when they can't be.
Brand capture went from "usually close" to "checked":
- Logo accuracy — sprite-referenced and icon logos resolve correctly, stored logo bytes are validated, and hidden menu images are no longer mistaken for logos.
- Colour palette — the logo's own colours lead the palette; roles are backfilled from two sources instead of guessed.
- Deeper brand facets — dark palette, font provenance, brand voice and style fingerprints, captured alongside the basics.
- A provider ladder — site scans try providers in order and report the real reason when they fail ("the page was blocked" rather than a vague error), with a retry that waits a beat.
- Manual edits win — a field you edit is locked to your value, and can be released back to the scanner deliberately.
A new structured-data audit end to end, plus one shared chat runtime and a truth panel showing exactly which tools an agent has.
Schema Markup Audit
- A complete new use case: crawl the rendered site, validate 13 structured-data types against a published rule table, and deliver a four-sheet workbook — coverage matrix, issues, and a generator for the markup that's missing.
- Sitemap-seeded sampling, with a human-in-the-loop pause you can answer and resume — the stream replays so you never lose context.
Agents
- Effective-tools truth panel — see exactly what an agent will have at runtime, with a test that stops the display drifting from reality.
- Built-in tools as the single control surface, with two-key build permissions and every chat surface declaring what it may call.
Everywhere
- Dashboard chat joined the shared chat runtime, so all surfaces stream, reason and ask identically.
- Answering a needs-input question resumes the same run instead of starting a new one.
Where a document came from, why a search matched, whether it really saved — and reports filed automatically under the right project.
A pass over the surface people live in:
- Provenance — every document shows where it came from, and its visibility is an explicit choice rather than an inherited default.
- Honest counts and empty states — the file count is real, and an empty view explains why it's empty instead of implying you have nothing.
- A Recent view and bulk delete for clearing up.
- Search that explains itself — results say what matched, so you know why you're looking at a file.
- Save confirmation — you're told your work saved, and told plainly when an upload's contents could not be kept.
- Filed automatically — an agent's report is stored under the proposal or project its activity belongs to, and indexed as it's written; notes can be re-indexed after editing.
- Places and contents browse in separate, resizable columns.
Ask in plain words — hide a tab, exclude a competitor, filter rows, rename a column, add a note — and the change is applied deterministically, on every format.
Audit deliverables now take edit requests in plain language:
- Say it, don't do it — "hide the annexure", "exclude BrandX", "show only rows under ₹2,000", "sort by rating", "rename SKUs to Products", "add a note for the client".
- Deterministic by design — the AI only chooses what to change from a closed vocabulary; the change itself is applied by code. Numbers are never touched, and the same ask always produces the same result.
- Every format — HTML reports and Excel workbooks alike, across all seven audit types.
- Edits persist — refresh or regenerate and your edits replay automatically.
- Requests outside the vocabulary are declined honestly and logged as real extension requests, so the vocabulary grows from demand.
Library
- The Reports grid was rebuilt with stat cards: each renderer publishes its own headline score and metrics, so a card shows real numbers at a glance, with refresh and download in reach.
One-click Amazon, Flipkart, Blinkit, Zepto, Instamart and PageSpeed audits — every number traceable to observed data.
The audit family landed — client-ready deliverables built from captured marketplace data:
- Amazon Listing & Store Audit — listing estate, pricing and discounting, A+ content coverage, review sentiment, per-category best-seller ranks and competitor comparison.
- Flipkart Listing & Store Audit — listing quality, content and image coverage, Assured coverage, search position and sponsored share.
- Quick-commerce audits (Blinkit, Zepto, Instamart) — pincode-level capture across multiple scan points per city: presence, availability, intra-city price variance, hero products and sponsored share, with capture time stated in local time.
- Website PageSpeed Audit — mobile and desktop Core Web Vitals as a client-ready workbook, with full-site budgets and locale-mirror collapsing.
What makes them trustworthy:
- Every figure is filled by code from a captured snapshot; the narrative is mechanically validated against those numbers, so a quote must be real and a claim must match its metric.
- Absent data is shown as absent — an unobserved value reads "not captured", never 0, and a scan point that wasn't serviceable is reported as a finding rather than silently skipped.
- Every recommendation carries an owner and a concrete first step.
- Category calibration knobs (discount risk, MRP flags) adapt the thresholds to the client's category.
A workspace manager with brands and drafts, plus priorities, true due dates, a calendar view and remembered board layouts.
Workspaces
- A Manage Workspaces view — resume an in-progress draft, group workspaces by brand, and create, rename or move brands inline.
- Archived workspaces stay out of the switcher; stale drafts are cleaned up; empty duplicates were de-duplicated and the actions that could destroy the wrong thing were removed.
Activities
- Priority and real due dates, with a board that reports status honestly.
- Calendar view as a fourth view, with drag-to-reschedule; clicking a date opens that day.
- Board layout is remembered between visits; a due date can be set straight from the card.
- Recurring runs fire in the schedule's own timezone rather than UTC.
- The run row shows the real run state and execution mode, with per-step input and output — it used to guess.
- Orphaned runs left by a restart are finalised automatically instead of hanging forever.
Adding a workspace opens with research instead of an intake form, and its questions are proper form fields — dates, amounts, pickers, even file answers.
Workspace setup was rebuilt around momentum:
- Research first — the up-front intake form is gone. Setup opens by researching the client on the web and comes back with findings, then asks about the gaps.
- Real form fields — questions render as native form controls: choices, numbers, amounts, date ranges, checkboxes, service-line pickers, and file answers you attach with a paperclip.
- It keeps going — a non-stop flow that stops interrogating and gets on with it; answers persist, so nothing is asked twice.
- A live brief — the client brief builds beside the conversation and is populated before you confirm.
- Resume anywhere — leave and return with the session and its full history intact.
- The progress bar and completion percentage were removed — they measured the form, not the work.
Skills
- Soft delete — a deleted skill stays deleted instead of resurrecting on the next reseed, with a Deleted view to restore from.
- Version history and the workflow bridge surfaced in the skill builder.
A codeless tool framework and an AI Tool Architect that drafts a working tool from an API's documentation.
Adding an integration stopped being an engineering ticket:
- Codeless tool framework — define a tool declaratively: endpoint, auth, inputs, and an output map with computed fields. Eleven existing tools were converted onto it.
- AI Tool Architect — point it at an API's documentation and it drafts the tool definition for you, catalog-aware so it doesn't duplicate what exists.
- New data tools — DataBlue scrape, crawl and map wrappers with a shared on-page SEO extractor; Firecrawl search, DeepSeek and Perplexity queries rewritten codeless with table fan-out.
- Honest failures — a draft that fails says why, inline, instead of a button that silently does nothing; the tool catalog refreshes itself so the skill builder never sees a stale list.
- Workbook layouts for XLSX deliverables, and styled workbook previews rendered page by page.