Examples gallery

The repo ships 14 runnable example crates under examples/, plus a configs/ gallery of YAML recipes. Each crate has its own main.rs and runs with cargo run -p <name>. Most run offline with a mock LLM — no API keys needed — unless noted.

These examples teach the library; they are not a runtime feature.

Coordination & planning

| Example | Demonstrates | Run | |---|---|---| | stigmergic-workflow | The EventLattice + depends_on DAG — running order emerges from pheromone signals crossing thresholds; no orchestrator decides it. | cargo run -p stigmergic-workflow | | skills-lattice | Event-driven Skills: one event fans out to every agent that reacts_to it (emits/reacts_to), distinct from a fixed DAG. | cargo run -p skills-lattice | | htn-planner | Symbolic HTN decomposition — compound tasks expand via methods; only unresolved frontiers reach the LLM. | cargo run -p htn-planner | | crash-recovery | Kill a multi-step workflow mid-run and resume from the checkpoint; completed steps are not re-run. | cargo run -p crash-recovery |

Memory & providers

| Example | Demonstrates | Run | |---|---|---| | memory-recall | Agent-managed core memory, semantic recall, and sleep-time consolidation (Tiers 3–4). Runs offline. | cargo run -p memory-recall | | multi-provider | Per-agent provider selection — a cheap local model for simple steps, a frontier model for the hard one, with a per-tier cost breakdown. | cargo run -p multi-provider |

Tools, protocols & integration

| Example | Demonstrates | Run | |---|---|---| | tool-hooks | Pre/post tool hooks that deny a path-traversal write, audit every call as JSON, and let the agent recover. | cargo run -p tool-hooks | | mcp-bridge | Call an external MCP tool over stdio through the real McpToolRegistry; plus exposing agents as an MCP server. | cargo run -p mcp-bridge | | a2a-server | Expose an agent over the A2A protocol (agent card + task endpoint) and call it from a client, in-process. | cargo run -p a2a-server | | sandbox-session | The rootless Podman sandbox for tool execution — threat model, config knobs, and a live integration test (needs Podman). | cargo run -p sandbox-session |

Autonomy & config

| Example | Demonstrates | Run | |---|---|---| | proactive-agents | Agents that fire on a schedule or on an event (here, reacting to AgentFailed), not on a user prompt. | cargo run -p proactive-agents | | configs/ | A gallery of minimal YAML configs (below). No Rust — validate and run with the CLI. | see Config gallery |

Foundations

The three foundational examples cover agent coordination, token budgets, and session/checkpoint memory:

| Example | Run | |---|---| | research-assistant | cargo run -p research-assistant | | code-reviewer | cargo run -p code-reviewer | | customer-support | cargo run -p customer-support |

Config gallery

examples/configs/ holds minimal, copy-pasteable axocoatl.yaml recipes — YAML

CLI only, no Rust. Each passes axocoatl validate, and each file’s header comment lists its prerequisites, run command, and expected output.

| File | Agents | Demonstrates | |---|---|---| | research-pipeline.yaml | 2 | depends_on workflow — summarizer waits on researcher | | feature-dev.yaml | 5 | Linear DAG (architect → planner → coder → reviewer → docs) | | incident-response.yaml | 3 | Skills (emits / reacts_to) + an ops agent | | local-only.yaml | 2 | Ollama, no API keys — data never leaves the box | | event-webhooks.yaml | 2 | Opt-in outbound event webhooks — the lattice POSTs signed events to a URL | | mcp-tools.yaml | 1 | A single MCP server (stdio transport) |

Validate or run any of them:

axocoatl validate examples/configs/research-pipeline.yaml
axocoatl dev -c examples/configs/research-pipeline.yaml

Most configs need a local Ollama (ollama serve && ollama pull llama3.2) — no cloud API key required. The mcp-tools recipe additionally needs npx on your PATH for the filesystem MCP server.