Engineering notes / Muhammad Kashif
Antigravity, Adobe, and Blender: Creative Tools That Survive Agents
Antigravity-class agent desktops, Claude driving Adobe tools inside chat, and Blender still sitting where serious 3D work goes — that trio forced me to ask which creative tools survive an agent in the loop without becomi…
Antigravity, Adobe, and Blender: Creative Tools That Survive Agents
By Muhammad Kashif · Draft · AI automation & cybersecurity builder, Pakistan
Antigravity-class agent desktops, Claude driving Adobe tools inside chat, and Blender still sitting where serious 3D work goes — that trio forced me to ask which creative tools survive an agent in the loop without becoming a liability. Prompting guides told me to define outcome, success criteria, constraints, and output shape, then let the model choose the process. I kept the guide. I also kept a human who can stop the process when Adobe, Blender, or a cloud desktop starts doing something I would not do sober.
Prompting guide as a seatbelt, not a personality
OpenAI's GPT-5.5-era prompting guide in the notes — outcome, success criteria, constraints, output, then freedom on process — is the brief template I now paste before creative agents touch brand files. Rigid step lists fight models; vague vibes fight clients. The middle is a contract: what good looks like, what is forbidden, what format lands in the folder.
I save that contract next to the export. When a stakeholder asks why a frame exists, I point at the criteria, not at a chat scrollback that evaporated with the session.
Adobe inside chat, Blender outside the chaos
Claude running Adobe tools inside chat is excellent for iteration speed on layouts and light edits. Blender stays the honesty machine for anything that needs real geometry, lighting discipline, or a file a studio can open without a vendor login. Agents can propose Blender steps; they do not get unsupervised write access to the master scene on day one.
Creative survival means the tool still works when the agent is wrong. If the only copy of the brand lives inside an agent transcript, the brand is already injured.
OpenClaw signing in with ChatGPT
OpenClaw using a ChatGPT Plus or Pro login without separate API credits lowers the barrier and muddies the metering story. Fine for personal experiments. For client creative ops I want billable metering, clear retention, and credentials that are not my primary chat login shared with a desktop agent.
Separate accounts, separate profiles, separate blast radii. Convenience that collapses identity boundaries is how one compromised session paints the whole week.
Enterprise side noise I actually track
China blocking Meta's Manus acquisition, Altman and Ive on a device, Google's Pentagon deal after dropping an older pledge, OpenAI and Microsoft rewriting exclusivity — these are dependency and ethics watch items, not creative tutorials. I track them so client advice about lock-in and policy risk is not naive.
I do not let geopolitics derail a Tuesday ad deadline. I do let it shape which vendors I am willing to bet a year's asset library on.
Agents that dream and self-check
Claude agents self-checking against a quality bar, plus overnight "dreaming" that reviews past work and cleans memory, sound like ops maturity. I want the quality bar in git where humans edit it. Overnight memory cleaning is welcome only if I control retention rules and can see what was forgotten.
Silent memory edits without a log are not hygiene. They are amnesia with PR.
Sheets, Teams, Chrome agents at the edge
ChatGPT inside Excel and Sheets, Perplexity inside Teams, Codex in Chrome — creative work increasingly sits beside spreadsheet truth and browser research. I keep creative write paths separate from finance sheets. An agent that can restyle a deck should not also be able to mutate revenue cells because someone enabled the same plugin pack everywhere.
Edge agents get edge scopes. Core numbers get fewer friends.
What "survives agents" means in a SOW
Source files in open formats. Brand tokens outside the chat. Prompt contracts filed. Human review on anything public. Credentials that are not personal daily drivers. If a tool cannot meet that list, it is a sketchpad, not a production creative system.
Antigravity-shaped desktops can host the sketchpad and the production lane — as long as the lanes do not share credentials or folders by default.
Friday creative audit
Exports offline? Masters not only in chat? Adobe and Blender credentials still scoped? Prompt contracts filed beside campaigns? Any overnight memory feature reviewed? If any answer is no, I fix it before the next campaign week starts pretending nothing happened.