ContextStack
Current roadmap

What ContextStack Is

View raw md
File: overview.md

ContextStack is building an owned organization of durable AI workers. Its Bots have missions, inboxes, commitments, scoped memory references, bounded authority, and performance history. A Bot remains the same accountable worker when the model, process, or machine performing one operation changes.

Product status: ContextStack is in active development and is not generally available. This page describes the finished product destination. The roadmap separates that destination from what has been demonstrated with evidence today.

The application uses the familiar grammar of threaded conversation, but chat is only a presentation and steering surface. Commitments, handoffs, approvals, evidence, and outcomes remain structured records so completed work can be reconstructed without reading a transcript.

The product model

The simplest way to understand the surface is:

  • Bot is who. A durable worker accepts responsibility across threads, workrooms, models, and time.
  • Workroom is where. Humans and Bots share commitments, decisions, evidence, and an outcome context.
  • Commitment thread is what. A conversation follows one Signal or Commitment, with a visible owner, state, next action, and outcome.
  • Run is how. Models, tools, sub-agents, execution cells, retries, and gates perform bounded work beneath the commitment.

Bot workspaces and shared workrooms are two views into one canonical work graph. Several Bots may contribute, but every accepted commitment has exactly one accountable Bot. Handoffs change ownership only when they are explicitly accepted.

Delegation is the interface

Work begins with a direct assignment or an external event normalized into a Signal. The organization routes that Signal to one accountable Bot according to mission and responsibility. People delegate an outcome; they do not have to draw an agent graph or select a workflow before useful work can begin.

Model choice is secondary by default. Gateway purchases the cheapest sufficient intelligence allowed by the work's privacy, quality, latency, and budget constraints. Every model purchase returns a receipt identifying what ran, under which policy, at what cost and latency, including any retry or escalation. The Bot's identity and responsibility do not change when the route does.

Governed execution beneath the thread

No-tool triage and planning can use receipted Gateway cognition under a standing, Kernel-governed Bot budget. Work that touches tools, artifacts, or external effects is commissioned as an order and performed inside a bounded run.

An execution cell receives only the capabilities resolved for that order. Those capabilities can include files, processes, a browser, or an ordinary application. When an application requires authentication, a broker grants the necessary session and credentials for a defined purpose and time. Credentials never become part of Bot identity, prompts, or memory.

The primary thread shows decisions, state changes, approvals, evidence, and outcomes. Model calls, tool activity, retries, and raw traces remain available as execution detail rather than dominating the conversation.

Autonomy stops at consequence

ContextStack separates permission to work from permission to create lasting effects:

  • Run-autonomy permits bounded work without a human watching each step.
  • Consequence authorization governs merge, deploy, publish, send, and spend separately, per lane and action class.

Revertible or compensable action classes may earn greater autonomy from recorded evidence. Actions without a reliable reversal remain explicitly approved or use an individually revocable, bounded, idempotent, and reconciled pre-authorization. See the autonomy ladder for the policy and current status.

Where to go next