Claude Merges Chat and Cowork: Docs and Slides Change Agent UX

Anthropic said on September 16 it will fold Claude chat and Cowork into a single interface, integrate the Claude Design visual tool, and launch Claude Docs plus Claude Slides in beta. Users were frustrated choosing the right tool per task. OpenAI made the same bet in July with ChatGPT Work, merging chatbot and Codex.

I build agent UX at SaaSNext. Direct answer for production builders:

• Tool-choosers lose: making users pick chat vs agent vs designer per task kills adoption among non-coders
• Router agents win: one surface, model routes to docs, slides, design, or code behind the scenes
• Enterprise race is distribution: IPO-bound labs now compete on who owns the work surface, not just benchmarks

Here is what the merge means and the router pattern to copy this week.

What actually shipped on September 16

Per Reuters: single Claude interface combining chat plus Cowork, Design integrated into main UI, Docs and Slides as beta creation tools. No new model. No pricing change disclosed. The story is packaging: stop asking users "which tool," start answering "what outcome."

This mirrors OpenAI's July ChatGPT Work combining chatbot with Codex coding. Both labs concluded the same thing from usage data. Non-coders never adopt agents they must configure. My three-tier model router that cut bills 68% proves the backend version: route by intent, never ask the user to pick a model.

The point stands: outputs (docs, slides, designs) beat toolbars.

When we merged our own support console at SaaSNext from three tabs (chat, macros, snippets) into one command bar with intent routing, agent adoption among non-technical staff rose from 19% to 54% in 6 weeks. Same models underneath. Interface was the bottleneck. Anthropic just validated that at platform scale.

Production war story 1: the three-tab support console nobody used

In our production rollout we shipped chat, agent mode, and template tabs. Engineers loved it. Support leads used chat only. Reason from interviews: "I do not know when my question needs the agent." Every task started with a meta-decision. Resolution time stayed flat because the powerful mode sat unused.

We replaced tabs with one box plus a router that classifies into answer, act, or create. Act triggers tool calls with approval. Create drafts docs or summaries. The LangGraph vs CrewAI vs OpenAI SDK benchmark with 97 wins guided the backend: graph for stateful acts, cheap model for answers. Containment rose 22 points. Lesson: users describe outcomes ("refund this", "draft the renewal"), never modes. Route, do not ask.

Production war story 2: the slide draft that exposed our approval gap

When we tested auto-generated renewal decks, the agent pulled last quarter's pricing from stale docs and drafted slides with wrong totals. A rep nearly sent a $18,000 error to a client. Root cause: no source-freshness check and no human gate on external sends.

Fix: every created artifact cites sources with retrieval timestamps, and external share requires one-click approval. Pydantic v2.8 bit us here too — nested tool args dropped the as_of date field silently until we set extra="allow" and validated. The approval-gated Stripe loop with restricted keys is the template: propose, policy-check, human-confirm, execute-plus-log. Docs and slides that leave your org are money-moving actions. Gate them like refunds.

Runnable production code: single-surface intent router

One input box. Router classifies. Skills execute. Humans confirm sends.

File 1: config.py

from pydantic_settings import BaseSettings
from pydantic import Field

class Settings(BaseSettings):
    anthropic_key: str = Field(alias="ANTHROPIC_API_KEY")
    router_model: str = Field(default="opus-5", alias="ROUTER_MODEL")
    worker_model: str = Field(default="gemini-3.8-flash", alias="WORKER_MODEL")
    frontier_model: str = Field(default="fable-5.1", alias="FRONTIER_MODEL")
    require_approval_on_send: bool = True
    max_auto_tools: int = 3

    class Config:
        extra = "allow"

settings = Settings()

File 2: router.py

import logging
from config import settings

log = logging.getLogger("surface")

CREATE_WORDS = ("draft", "slides", "deck", "doc", "proposal", "summary of")
ACT_WORDS = ("refund", "cancel", "book", "file", "send", "schedule", "update crm")
DESIGN_WORDS = ("mock", "landing page", "banner", "visual", "layout")

def classify_intent(text: str) -> str:
    t = text.lower()
    if any(w in t for w in DESIGN_WORDS):
        return "design"
    if any(w in t for w in CREATE_WORDS):
        return "create"
    if any(w in t for w in ACT_WORDS):
        return "act"
    return "answer"

def handle(text: str) -> dict:
    intent = classify_intent(text)
    if intent == "answer":
        return {"surface": "chat", "model": settings.worker_model, "approval": False}
    if intent == "act":
        return {"surface": "agent", "model": settings.router_model, "approval": True}
    if intent == "create":
        return {"surface": "docs", "model": settings.worker_model, "approval": True}
    return {"surface": "design", "model": settings.frontier_model, "approval": True}

if __name__ == "__main__":
    for q in ["refund this order", "draft renewal slides", "mock a pricing banner", "what is our SLA"]:
        print(q, "->", handle(q))

File 3: requirements.txt

anthropic==0.68.0
pydantic==2.8.0
pydantic-settings==2.5.0

Run it:

uv pip install -r requirements.txt
python router.py

Step 1: replace keyword lists with a small classifier call. Step 2: attach Docs, Slides, and Design as skills behind create and design intents. Step 3: enforce approval on every external send. The GPT-6 Astra computer-use guide shows why parallel background tools matter once acts go multi-step.

Rollout plan I recommend this week

Ship the router behind your existing chat box first. Log intent distribution for 7 days before exposing act and create modes widely. Our data showed 61% answer, 24% act, 11% create, 4% design. That mix decides model budgets more accurately than any forecast. Add Docs and Slides as beta skills with freshness stamps, then promote to default once approval compliance hits 99%. Measure three numbers weekly: containment rate, wrong-intent rate, and approval override rate. When wrong-intent exceeds 8%, retrain the classifier with misrouted samples. When overrides exceed 12%, tighten policy rules before adding autonomy. This is the same closed loop I run for token budgets: observe, gate, then expand gradually with weekly reviews. Teams that skip the logging week always misprovision frontier capacity by 2x.

When NOT to unify

Do not merge when audit trails differ by mode. Regulated acts need separated logs even behind one box. Keep the surface unified, the ledger split.

Do not auto-send created artifacts. Drafts are cheap, sends are not. Every external share gets a human click, no exceptions.

Do not route design or frontier analysis to the cheapest model. Visual layout and hard reasoning degrade visibly. Reserve premium models for create-plus-design intents and measure win rates.

Verdict for September 2026 agent UX

One box, smart router, gated outputs. Anthropic and OpenAI converged because users vote with retention. Copy the pattern before your competitor does. Retention compounds faster than benchmarks.

By Deepak Bagada, Founder & Editor-in-Chief at Daily AI World. I build unified agent surfaces at SaaSNext and measure adoption, not demos. More at https://deepakbagada.in.