Agents & policies

Your delegate runs as a single agent with a focused set of tools. When it needs information beyond what it already knows, it reaches for a tool — most commonly web_search for up-to-date facts — and weaves the result into its reply. You just ask; it picks the right tool.

What it can do

The base agent has a curated set of capabilities: messaging (SMS, email), notes, persistent memory, GitHub, calendar (via MCP), a browser-driven Copilot for interactive tasks, and web_search for fresh information from the web. If something takes a little longer, that’s a tool call running behind the scenes — on phone calls a gentle hold-music pulse keeps you company while the lookup completes.

Capabilities are managed automatically. To customise which tools the agent can use, see the Technical details below.

Technical details

Single base agent

Base agent — configured in src/agentConfigs/baseAgentConfig.ts. Handles every turn. Calls tools (local handlers, MCP tools, or web_search) as needed.
web_search tool — local function tool (src/tools/handlers/web-search.ts) that wraps a single OpenAI Responses API call with the builtin web_search tool enabled. The wrapper exists because the Realtime API (voice) only accepts function tools, not OpenAI builtin tool types — wrapping makes web search available on text, voice, and phone uniformly.
Agent registry — src/agentConfigs/index.ts. Each agent has an id, instructions (system prompt), model, and tools (resolved at runtime via policies).

Policies

Per-agent tool access is controlled by runtime-data/agent-policies.json. A policy is an allowlist combining:

Tags — e.g. base-default grants every tool tagged that way
Names — explicit allow of individual tool names

To customise, edit runtime-data/agent-policies.json directly and restart, or use the settings UI. At startup the log will confirm: [registry] Loaded persisted policies for 1 agent(s).

Turn flow

user message ──▶ base agent ──▶ answer (direct)
                         │
                         └▶ tool call (web_search / local.* / mcp.*)
                               └▶ tool result fed back ──▶ final answer

The standard function-call loop runs up to 8 iterations, allowing the agent to chain tools (e.g. web_search followed by create_note and send_sms) before producing the final reply.

Inner-context flow

The Inner Context Plane extends activation beyond user turns while preserving one base agent:

memory / task / timer / subprocessor
              │
              ▼
        durable inner signal
              │
              ▼
        attention + context ──▶ base agent ──▶ existing tools

Subprocessors contribute awareness rather than becoming additional agents. Tool policy remains attached to the base agent, and prompts plus standing memories determine whether a signal warrants action.