Roadmap
All planned work across the project, organized by area. The end-to-end agent loop is the
highest priority — everything else is sequenced around making that work reliably first.
completed
in progress
next
planned
backlog
78% complete
Core System
Agent execution loop, model routing, config schema, context engineering, memory, cost tracking, and evaluation infrastructure.
| ID | Task | Status | |
|---|---|---|---|
| task-010 | Go module and project skeleton F03F02 | completed | |
| task-011 | Agent config schema (TOML) and loader F05 | completed | |
| task-040 | Model router — Anthropic + OpenAI-compat + fallback chain F10F06 | completed | |
| task-041 | Core agent execution loop with NextStep and Hooks F07F10 | completed | |
| task-042 | Context engineering layer — ContextEngine F06F05 | completed | |
| task-043 | Tool executor — parallel dispatch, partial failure, timeouts F10F11 | completed | |
| task-044 | Security layer — prompt injection detection + A2A input validation F05F06 | completed | |
| task-045 | Model fallback chain and cost circuit breaker F03F06 | completed | |
| task-022 | 4-layer memory architecture F07 | completed | |
| task-030 | AI-assisted config generation prompt and engine F02F05 | completed | |
| task-031 | Onboarding interview agent — ai init F02F04 | completed | |
| task-046 | Prompt caching optimization F06F03 | completed | |
| task-047 | Tool context management for large tool lists F06F04 | completed | |
| task-083 | Durable session store — FileSessionStore + crash recovery F07 | completed | |
| task-085 | Cost observability and budget reporting F11F06 | completed | |
| task-156 | Human-in-the-loop: built-in human_approval MCP tool F07 | medium | completed |
| task-157 | Explicit control flow: intent router and tool-call guards F05F06 | low | completed |
Validation spikes (completed)
| task-001 | Go agent framework evaluation F03F09 | completed | |
| task-003 | Anthropic Go SDK — streaming + tool-use validation F06 | completed | |
| task-005 | MCP Go library selection F09 | completed | |
| task-006 | MCP dual-role PoC — agent as client AND server F09F10 | completed | |
| task-007 | Context engineering validation harness F06 | completed | |
| task-002 | Kuzu Go bindings validation F08 | completed | |
| task-004 | mcp-toolbox Neo4j source evaluation F09 | completed | |
| task-160 | Prompt injection hardening in agent loop F05F06 | high | completed |
| task-161 | Reversibility taxonomy + default-approve-required for irreversible tool calls F07 | high | completed |
| task-162 | Trust hierarchy: principal tiers + trust-level enforcement in agent loop F05 | high | completed |
| task-163 | Platform safety floor: non-overridable hardcoded behaviours F11 | medium | completed |
| task-164 | Honesty guidelines: uncertainty language, capability statements, AI identity F05 | medium | completed |
| task-170 | Wire agent.Agent.Run as real TaskRunner in ai serve F07F09 | critical | completed |
| task-172 | Config-to-agent wiring verification: context thresholds and tool call Index sort F06F05 | high | completed |
| task-173 | cmd_init test coverage F02 | medium | completed |
| task-270 | GenerateToolboxConfig — read_cypher + write_cypher opt-in tools F10F05 | high | planned |
| task-272 | write_cypher — wire as R_DESTROY in generated agent.toml and built-in patterns F07F05 | high | planned |
| task-274 | Integration test — read_cypher executes live Neo4j query via mcp-toolbox F11 | medium | planned |
| task-281 | Dynamic tool descriptions — inject agent name + purpose into agent_run/agent_stream F02F01 | high | planned |
| task-290 | TokenValidator interface + auth middleware for MCP and A2A servers | critical | planned |
| task-291 | RFC 9728 Protected Resource Metadata endpoint | high | planned |
| task-292 | JWT validation with JWKS auto-discovery and key rotation | high | planned |
| task-293 | A2A AgentCard OAuth2 SecurityScheme + scope enforcement | medium | planned |
| task-294 | Auth configuration in agent.toml + deploy.auth schema F02 | high | planned |
| task-298 | E2E auth integration test: MCP client → OAuth → tool call | high | planned |
CLI
The ai command and its subcommands — the primary interface for developers
running, configuring, and serving agents locally.
| ID | Task | Status | |
|---|---|---|---|
| task-012 | CLI framework and top-level command structure F01F03F12 | completed | |
| task-031 | ai init — onboarding interview agent F02F04 | completed | |
| task-051 | ai serve and ai run — serve and single-run modes F01F02 | completed | |
| task-054 | Curl-based install script and binary distribution F02F03 | completed | |
| task-135 | Version in help output and ai with no arguments F01F12 | completed | |
| task-052 | ai show — display agent config with syntax highlighting F05F04 | completed | |
| task-053 | ai web — local developer console (chat + trace + cost) F11F04 | low | completed |
| task-136 | Global --quiet / -q flag for silent mode F01F12 | completed | |
| task-159 | ai init — progressive elicitation with agent type templates F02F04 | medium | completed |
| task-153 | ai show --diagram — ASCII architecture diagram via D2 F11F04 | low | completed |
| task-155 | --plain flag — screen reader and pipeline-safe output F12 | low | completed |
| task-165 | ai init validation: warn when generated config violates our own principles F04F10 | low | completed |
| task-180 | ai init: readline support for cursor/arrow key navigation F11F01F02 | medium | completed |
| task-181 | ai init: fix model name suggestions and add normalization F04F01 | high | completed |
| task-182 | ai init: investigate and fix crash after model name entry F04 | medium | completed |
| task-183 | ai init: fix --agent → --config flag name in 'Next step' output F04 | low | completed |
| task-184 | ai init: --credentials flag for Aura/Sandbox credential file loading F01 | low | completed |
| task-185 | NDJSON done event: include cost_usd, input_tokens, output_tokens F04F11 | medium | completed |
| task-186 | ai init: demo database catalog — gather and add known demo databases F02F03 | medium | completed |
| task-188 | ai serve: auto-detect and start mcp-toolbox sidecar without explicit flag F02F11 | high | completed |
| task-193 | ai graph seed — seed local Kuzu DB from named dataset or Cypher file F08F02F01 | medium | planned |
| task-191 | ai init — local-first Kuzu path (option 5 in URI preset menu) F02F08F04 | high | planned |
| task-194 | ai init — AI-generated domain seed data for local Kuzu F02F08F04 | medium | planned |
| task-208 | Extend --credentials to parse Supabase, PlanetScale, Neon, MongoDB, Redis URLs F02F03 | medium | planned |
| task-240 | ai init — database type selector menu (replaces URI preset menu) F02F04F01 | high | planned |
| task-248 | Extend --credentials to parse Supabase, PlanetScale, Neon, MongoDB, Redis URLs F02F03 | medium | planned |
| task-175 | Shell completion: ai completion bash|zsh|fish — static completions F04F01 | medium | completed |
| task-176 | Shell completion: dynamic completions — config files, skills, task IDs F04F01 | low | completed |
| task-177 | Coding agent UX: --output flag + structured errors + exit codes on all commands F11F01F04 | high | completed |
| task-178 | Coding agent UX: --non-interactive, stdin piping, AI_ env var overrides F02F01F04 | high | completed |
| task-179 | Coding agent UX: ai status, --dry-run, --timeout, and completion self-doc F11F07F02 | medium | completed |
| task-250 | ai show — opinionated default view (identity + truncated prompt + non-defaults) F01F02 | high | completed |
| task-251 | ai show — --full flag (preserve current full TOML dump behavior) F01 | high | completed |
| task-252 | ai show — --tools flag (parse toolbox YAML, show sources + tools + env var status) F02F03 | high | completed |
| task-253 | ai show — --section flag (show single named config section) F01 | medium | completed |
| task-254 | ai show — --filter flag (interactive fzf / built-in line filter) F01F08 | medium | completed |
| task-255 | ai show — --grep flag (non-interactive regex line filter) F01 | low | completed |
| task-271 | ai init — offer generic Cypher fallback tools in onboarding interview F02F04F05 | high | planned |
| task-273 | ai show --tools — annotate fallback tools with visual separator and warning F11F01 | medium | planned |
| task-280 | ai mcp — MCP stdio server command F02F01 | high | planned |
| task-282 | ai show --mcp-config [target] — print ready-to-paste coding agent config snippets F05F11 | high | planned |
| task-283 | ai init — write MCP config files for Claude Code and Cursor at end of onboarding F02F01F11 | medium | planned |
| task-284 | ai init — write MCP config files for Claude Code and Cursor at end of onboarding F02F04 | high | planned |
| task-285 | ai show — MCP tool identity preview section F11F05 | low | planned |
Integrations
MCP client and server, mcp-toolbox sidecar, graph backends (Neo4j, Kuzu), and Python sidecars for GraphRAG, graph construction, memory, and eval.
| ID | Task | Status | |
|---|---|---|---|
| task-034 | MCP client — agent connects to external MCP servers F09 | completed | |
| task-060 | Skills system F10F05 | completed | |
| task-070 | Python sidecar service manager (Go) F10F08 | completed | |
| task-020 | Neo4j connection and schema introspection F08F09 | completed | |
| task-032 | mcp-toolbox config generation and sidecar manager (ai sidecar mcp-toolbox) F09F08 | completed | |
| task-033 | MCP server — agent exposes its tools over MCP F09 | completed | |
| task-071 | GraphRAG retrieval sidecar (Python, port 8091) F08F10 | completed | |
| task-072 | Graph construction sidecar (Python, port 8090) F08F10 | completed | |
| task-073 | Agent memory sidecar (Python, port 8092) F07F08 | completed | |
| task-021 | Kuzu local graph backend F08 | completed | |
| task-158 | Schema-first tool definitions: Go structs as tool contracts F09 | low | completed |
| task-171 | MCP client reconnect backoff implementation F03F11 | high | completed |
| task-174 | mcpserver agent_stream: true SSE relay from agent execution F09F11 | medium | completed |
| task-190 | Add [agent.graph] config section (backend, kuzu_path, seed_dataset) F08F02F04 | high | planned |
| task-192 | Embed seed datasets: recommendations, northwind, countries F08F02F03 | medium | planned |
| task-195 | Wire Kuzu backend into ai serve as runtime graph query target F08F02F11 | high | planned |
| task-206 | Config generator: Redis / Valkey F03F02 | low | planned |
| task-207 | SQL schema introspection for ai init (postgres, mysql, sqlite) F02F04F11 | medium | planned |
| task-209 | SQL seed data for DuckDB and SQLite (recommendations, northwind, countries) F08F02F03 | medium | planned |
| task-241 | Config generator: PostgreSQL (Supabase, Neon, plain postgres) F02F03F08 | high | planned |
| task-242 | Config generator: MySQL (PlanetScale, plain mysql) F02F03 | high | planned |
| task-243 | Config generator: SQLite (local embedded relational) F08F02F03 | medium | planned |
| task-244 | Config generator: DuckDB (local) and MotherDuck (cloud) F08F02F03 | medium | planned |
| task-245 | Config generator: MongoDB Atlas (document store) F02F04 | medium | planned |
| task-246 | Config generator: Redis / Valkey F03F02 | low | planned |
| task-247 | SQL schema introspection for ai init (postgres, mysql, sqlite) F02F04F11 | medium | planned |
| task-249 | SQL seed data for DuckDB and SQLite (recommendations, northwind, countries) F08F02F03 | medium | planned |
Web UI
Local developer console — chat, trace, cost counter, eval integration, graph explorer. See web.html for design mockups. Depends on the agent execution loop (task-041) and REST API (task-091).
| ID | Task | Status | |
|---|---|---|---|
| task-091 | REST management API — /api/* endpoints F09F04 | completed | |
| task-090 | Web UI shell — routing, theme, agent list, config editor F04F11 | low | completed |
| task-053 | ai web v1 — chat + ASCII trace + cost counter + eval F11F04 | low | completed |
| task-142 | Graph explorer tier 1 — schema view (read-only) F11F08 | low | completed |
| task-141 | Full eval tab — run evals, dataset management, diffs F11F04 | low | completed |
| task-143 | Graph explorer tier 2 — Cypher playground F11F08 | low | completed |
| task-147 | Session persistence — JSON/YAML in .ai/ folder F07 | low | completed |
| task-152 | Live streaming trace — spans appear in real time F11 | low | completed |
| task-148 | Session tab switcher and history browser F07F04 | low | completed |
| task-145 | Web accessibility — WCAG 2.1 AA F12 | low | completed |
| task-154 | D2-powered trace diagram — optional visual mode F11 | low | completed |
| task-146 | TUI mode — ai web --tui / ai tui via bubbletea F01F12 | low | completed |
| task-150 | System prompt diff on agent.toml reload F05 | low | completed |
| task-151 | Span replay — re-run any LLM or tool call in isolation F11F07 | low | completed |
| task-144 | Graph explorer tier 3 — visual canvas (Neo4j NVL) F11 | low | completed |
| task-149 | Team sharing — auth, per-user session isolation F09 | low | completed |
| task-265 | ai web — command palette keyboard navigation + fuzzy search F04F01 | medium | completed |
| task-266 | ai web — auto-start ai serve when backend is not running F02F04 | high | completed |
| task-260 | ai web — graceful port conflict handling + reload signal F02F04 | medium | completed |
| task-261 | ai web — unified Data tab with source switcher and generic schema view F04F11F08 | medium | completed |
| task-262 | ai web — SQL query playground (postgres, mysql, sqlite, duckdb) F11F04 | medium | completed |
| task-263 | ai web — MongoDB document browser F11F04 | low | planned |
| task-264 | ai web — Redis / Valkey key browser F11F04 | low | planned |
Deployment & Production
A2A protocol server, OpenTelemetry instrumentation, evaluation framework, Fly.io and Cloud Run deploy targets, distribution, and integration test suites.
| ID | Task | Status | |
|---|---|---|---|
| task-050 | A2A protocol server with auth and task ownership F09F10 | completed | |
| task-055 | A2A auth middleware and rate limiting F09 | completed | |
| task-080 | OpenTelemetry instrumentation F11 | completed | |
| task-081 | Agent execution trace graph writer F11 | completed | |
| task-082 | Evaluation sidecar (Python, port 8093) F11F10 | completed | |
| task-084 | Structural evaluation — no-LLM, CI-runnable F11 | completed | |
| task-110 | Distribution — go install, brew tap, GitHub releases F02F03 | completed | |
| task-120 | Integration tests — end-to-end agent bootstrap F02F11 | completed | |
| task-121 | Integration tests — graph construction and GraphRAG retrieval F08F11 | completed | |
| task-122 | Acceptance tests — mcp-toolbox + Neo4j article and codelab flows F09F11 | completed | |
| task-123 | Integration tests — LLM provider and model router (Anthropic + OpenAI) F10F11 | completed | |
| task-124 | Integration tests — Neo4j and Aura database connectivity F08F11 | completed | |
| task-100 | ai deploy — Fly.io container deployment F02F08 | completed | |
| task-101 | Cloudflare domain management and edge routing F02 | completed | |
| task-102 | ai deploy — Cloud Run for graph construction jobs F08F10 | completed | |
| task-103 | Spike — WASM/WASI feasibility (Cloudflare Workers + agent binary) F03F08 | completed | |
| task-187 | CLI: add make install target and ai --version flag F02F01 | medium | completed |
| task-295 | Cloudflare Worker OAuth 2.1 auth proxy with DCR + CIMD F02 | medium | planned |
| task-296 | ai deploy --target cloudflare command F02 | medium | planned |
| task-297 | ai deploy --target cloudrun with IAP auth F02 | low | planned |
Demo Series
Five end-to-end walkthroughs — movie recommendations, company intelligence,
clinical knowledge graph, multi-agent A2A pipeline, and cloud deployment.
Includes fixture configs, validation scripts, and an interactive demo page at
agent-intelligence.ai/demos.
| ID | Task | Status | |
|---|---|---|---|
| task-200 | Demos: Write demo-01 movie recommendations script F02F04 | high | completed |
| task-201 | Demos: Write demo-02 company intelligence + GraphRAG script F09F11 | high | completed |
| task-202 | Demos: Write demo-03 clinical knowledge graph script F10F11 | high | completed |
| task-203 | Demos: Write demo-04 multi-agent A2A pipeline script F07F09 | high | completed |
| task-204 | Demos: Write demo-05 deploy + Claude Desktop script F02F09 | high | completed |
| task-205 | Demos: Build docs/demos.html interactive demo selector page F01F04 | high | completed |
| task-210 | Demos: Create fixture agent.toml + toolbox.yaml for each demo F02 | high | completed |
| task-220 | Demos: Implement A2A client call from agent (agent → agent delegation) F09F10 | high | completed |
| task-221 | Demos: Add --stream flag to ai run for SSE output F01F11 | high | completed |
| task-222 | Demos: Add --endpoint and --token flags to ai run F01F09 | high | completed |
| task-223 | Demos: Package and host fixture tarballs for quick-start F02 | medium | completed |
| task-225 | Demos: Document Claude Desktop MCP integration in docs/api.html F01F09 | medium | completed |
| task-226 | Demos: Document ai run --endpoint, --token, --stream flags in CLI reference F01F04 | medium | completed |
| task-227 | Demos: Add multi-sidecar architecture section to docs/architecture.html F10F11 | medium | completed |
| task-228 | Demos: Add multi-agent A2A topology section to docs/api.html F07F09 | medium | completed |
| task-230 | Demos: Verify all demo Cypher queries against live Neo4j demo databases F11 | high | completed |
| task-231 | Demos: Create end-to-end validation script for demo 01 F11 | medium | completed |
| task-232 | Demos: Add Demo Series section and link to docs/roadmap.html F02 | low | completed |
Docs & Landing Page
Public site at agent-intelligence.ai — landing page, docs pages, content
compliance, analytics, and deployment to Cloudflare Pages.
| Task | Status | |
|---|---|---|
| Landing page — above-the-fold layout, hero, feature strip | completed | |
| Terminal animation engine and screenplay | completed | |
CLI reference page — docs/cli.html | completed | |
Config reference page — docs/config.html | completed | |
API reference page — docs/api.html | completed | |
| Architecture page with ASCII diagrams (D2 pipeline) | completed | |
Web UI design page — docs/web.html with TUI mockups | completed | |
Roadmap page — docs/roadmap.html | completed | |
Demos page — docs/demos.html interactive demo selector | completed | |
docs/llms.txt for LLM navigation | completed | |
| Copy buttons and curl install block on landing page | completed | |
| GitHub star badge / widget on landing page | planned | |
| Inline email capture form with backend endpoint | planned | |
| Cloudflare Web Analytics script tag | planned | |
| Configure Cloudflare Pages deployment | planned | |
| Full accessibility pass — WCAG 2.1 AA | planned | |
| Performance audit and asset weight optimization | planned | |
| Content compliance — remove internal tech framing from copy | planned | |
| Cross-browser validation and smoke test | planned |
Sandbox & Code Execution
Secure multi-tier code execution: JavaScript via QJS + wazero (tier 1), Python via CPython WASM (tier 2a) or Firecracker (tier 2b), mini-Linux via Firecracker + nsjail fallback (tier 3).
| ID | Task | Status | |
|---|---|---|---|
| task-130 | Design secure code execution sandbox subsystem F10 | completed | |
| task-131 | JavaScript sandbox — QJS + wazero, tier 1 F10F08 | completed | |
| task-132 | Python sandbox — CPython WASM (tier 2a) and Firecracker (tier 2b) F10F08 | completed | |
| task-133 | Mini-Linux sandbox — Firecracker + nsjail fallback, tier 3 F10F08 | completed | |
| task-134 | Sandbox MCP tools and code-execution skill F09F10 | completed | |
| task-132b | Python Firecracker sandbox — Tier 2b (follow-on to task-132) F10F08 | planned | |
| task-132b | Python Firecracker sandbox — Tier 2b (follow-on to task-132) F10F08 | planned | |
| task-132b | Python Firecracker sandbox — Tier 2b (follow-on to task-132) F10F08 | planned |