kallipai is a tagma-centric agent runtime aimed at cross-project, multi-agent coordination - early stage, not yet ready for production use. This document captures where the project is headed; see architecture.md for how it works today.
Goal: a local agent team that meets basic agent needs, like coding - built on agentic design.
- Agentic agent management - agents spawn and manage other agents; the team organizes itself, not driven by a human operator top-down
- Non-blocking permission control - risky actions go through async approval (deferred), so the agent keeps working while approval happens elsewhere
- Inter-agent communication - async messaging between agents
- Self-managed agentic context - each agent controls its own context (pin / evict), not silently summarized by the system
- Team-evolving skill system - skills accumulate from the team's experience and evolve with it, shared across the team
- Misuse-resistant toolset - tools shaped so agents rarely misuse them
- Production hardening - reliability, persistence, observability, cost/usage tracking, test coverage
Goal: agent teams that deploy and scale fast, manageable on the go via mobile and cloud.
- Deployment & scaling - quickly spin up and scale out agent teams
- Mobile app - manage agent teams from your mobile device
- Cloud control - bridge mobile clients to teams across instances
Goal: a zero-employee company - run entirely by agents, serving its own customers and taking on work as one entity.
- Multi-team composition - the company is built from multiple coordinating teams (the deployable units from P2)
- Customer-facing - it has its own customers, takes on tasks and projects, and delivers outcomes
- Fully autonomous - run end-to-end by agents with no human employees; humans are customers, not staff
- Organizational memory - roles, hierarchy, and knowledge persist as the company's, beyond any single agent
- Sandbox breadth - the dynamic FS scoping and per-directory write-mutex core (a read/write lock so only one agent may hold write access to a directory at a time, for multi-agent concurrency safety) has landed, enforced per spawned shell via landlock on Linux. Still-open isolation axes: cgroup resource limits, network egress control, overlayfs copy-on-write staging, and a secret-proxy tool.
- Tool-call metrics for harness refinement - tool-call outcomes (command exit codes and the misuse they imply, success rates, failure patterns) are a feedback signal, not a model ranking: with the model held fixed, they show which of our tool and CLI designs lead the agent into misuse, guiding continuous refinement toward rarer misuse