What shipped.

Coordination and memory.

Added

• Coordinator role: an agent with role: coordinator decomposes a goal into subtasks, assigns each to a worker by a capability and budget auction, runs them in parallel, and synthesizes the result. Symbolic HTN planning when methods are provided, otherwise LLM decomposition. Resumable from the last checkpointed subtask.
• Agent-editable core memory (the MemGPT model): small blocks the agent curates as it works, rendered into its context every turn. Agent-driven recall (recall_search, recall_timeframe) and a background sleep-time pass that promotes durable facts up into the core blocks.
• Automatic context compaction: long conversations summarize toward the model window instead of being snipped, so early context survives.
• OpenAI-compatible servers: point the openai provider at LM Studio, MLX, or vLLM via base_url, with the agent's model sent per request. Stdio MCP servers now receive their env vars.

Changed

• Tier 3 is now agent-editable core memory; the old shared long-term fact store is retired.
• overflow_policy is strictly a spend cap (abort default, or warn); context compaction is automatic and independent of the budget.

The cockpit.

Added

• Variants: fan one prompt into N parallel attempts, each a real agent in its own git worktree, and keep the one you like.
• A conversation-forward session cockpit: an N-surface layout (files, activity, browser, terminal, agent graph) that starts as just the conversation. Agent edits land as change cards; a running dev server lands as a preview card.
• Unified conversation UI across Chat and Sessions; persisted session transcripts with Retry and Rewind.
• Git-native sessions: a live Source Control pane (status, diff, commit, discard, branch switch) inside the cockpit.
• Multi-turn tool-calling verified across all six providers.

The runtime that runs.

Added

• Stigmergic multi-agent coordination via EventLattice — agents activate themselves off accumulated signals, with no central orchestrator.
• Workflow execution — CLI, HTTP, and IPC paths. Entry agents activate directly; downstream agents cascade via TaskCompleted events.
• Per-agent token budgets, four-tier persistent memory with neural Tier 4, and checkpointed agent state.
• Directory sessions with podman-sandboxed bind-mounts, full PTY terminal support, and resumable state.
• MCP integration: bundled server catalog, install flow, permission gate, and a Tools view in the dashboard.
• Provider matrix: Ollama, OpenAI, Anthropic, Mistral, Gemini, OpenRouter. Per-agent selection.
• Always-on system service (systemd / launchd) via axocoatl service install.
• Interactive axocoatl onboard + axocoatl doctor.
• Dashboard: Sessions cockpit, Studio lattice canvas, Automations editor, Chat, Files, Skills, MCP, Agents.
• 340+ tests passing across the workspace.