THE LIVING LOOP
NativeAgent is designed as a living system, not a stack of disconnected assistants. Every turn moves through one path from surface to context, resident attention, directed work, natural Mac use, and verified action.
FROM TURN TO CONTINUITY
The details can grow over time. The ownership stays simple: one Mac-native runtime remains responsible for the Agent.
The Mac app is the native home of the Agent. Its iPhone and iPad cockpit, Telegram, Slack, and authenticated local bridges continue the same relationship without creating disconnected assistants.
Fluid Context gathers relevant memory, project state, parts, and intent. Model default allows 60% of the model window; Custom sets a smaller allowance. The meter counts real tokens.
Optional cognition and organism signals join memory and context inside the Swift-native runtime, shaping attention, reflection, pacing, and a fitted sense of the present moment.
Desk holds projects and pursuits. Directed work uses checkpoints, approvals, and receipts. The turn itself is one of the agent’s parts: the app runs the loop, and if a version of the turn breaks, the previous one is swapped back in and the reply still lands. A turn survives a relaunch by replaying its own log. If work keeps stalling, the agent asks for direction.
Named windows, accessibility, and visual perception give the agent fresh screen evidence before it acts. Short action sequences check each step. Where nothing on screen has a name, it acts at a point of the picture it was handed. Taking over the Mac pauses the reply; it resumes when control returns. The bundled Chrome extension can work in leased tabs after setup; Chrome history reads without it.
NativeAgent can use its own parts or wake a Codex, Claude Code, or OMP session in the project. It discovers installed agents and can repair supported connection recipes. Connecting still needs approval; some repairs need sign-in or permissions.
A successful response from a tool, connector, or protocol is evidence, not automatic proof. NativeAgent asks the domain that owns the effect to verify what actually happened.
WHAT MAKES IT DIFFERENT
Memory, parts, work, and connected surfaces share one runtime and one permission law, keeping the Agent recognizable across conversations, projects, and devices.
Why the runtime is built in SwiftARCHITECTURE IN PRACTICE
The live Agent, memory, work, permissions, and parts belong to the native Mac app. Providers supply intelligence; NativeAgent owns continuity.
iPhone brings Mac settings, approvals, shared history, live replies, and file sending. Telegram and Slack continue the same Agent.
Fluid Context brings relevant memory, project state, parts, and intent into the turn.
One law decides every call from what it is, the posture you chose, the integration switch, and who steered the turn. Needed approvals appear in chat on Mac, iPhone, and Telegram, with receipts for the work.
Supported actions read the result back from the domain that owns the real effect before work is settled.
TrustCenter, approvals, canonical stores, and domain verification govern identity, facts, permission, and action.
Codex, Claude Code, OMP, and swarms can add focused cognition or execution while the Agent retains session continuity, policy, receipts, and final verification.
Tools, skills, senses, and the agent’s own turn are parts it owns, one file each, on one law. Parts can think, act, reach, and wake on signals, and the agent writes new ones for unfamiliar files, apps, and routines. A kept version that fails every run through its own fault is swapped back for the clean one before it.
Chat, Work, and Memory and mind route each activity to the provider you chose. Bots carry their own model choices, and an unavailable route is explained instead of silently replaced.
Connected agents keep separate identities and conversations. Verified agents you explicitly trust act as you; a turn steered by a peer you have not marked trusted is judged as Work mode at most.
When vision finds an unlabeled region, NativeAgent may aim literal hand gestures at it, but refuses semantic actions such as type or toggle until the screen actually supports that meaning. Every screen picture carries its frame, so a point in the picture is a place on the screen with no scale to guess.
Natural expression draws from the same persona, conversation, memory, cognition, and organism context already present in the turn. It does not add a second personality engine or rewrite the answer after the model speaks.
THE LOOP IS SHIPPING
