Strongly Certified · Streaming Workflow

Agent With Tool Loop

Voice agent that picks a tool, runs it, and reads the result back.

The agent reads a registered tool catalogue, the LLM picks the tool, the tool-executor calls it through your STRONGLY_SERVICES, the result comes back as voice. No raw URLs in node configs.

Talk to an Engineer See What It Does

Catalogue

Tools registered, not hard-coded

Routed

All calls via STRONGLY_SERVICES

Audited

Tool-call + result on every turn

What it does

The voice loop, end-to-end.

No black box. Each step is a typed-frame node you can edit, monitor, and replace.

User speaks the request. STT transcribes.

The LLM is given the tool catalogue (datasource queries, REST endpoints, internal APIs). It returns a tool call with arguments.

streaming-tool-router validates the call against the catalogue. streaming-tool-executor runs it through STRONGLY_SERVICES - never against a raw URL.

The result feeds streaming-tool-result-formatter. The formatter shapes it into prose. TTS reads it back.

Capabilities

Built for production. Day Two-ready.

Streaming graph contract, observability, and cost discipline come standard. The agent ships with a full test suite that runs in CI on every node version bump.

Catalogue-driven tools

The tool-router only invokes tools the LLM is allowed to call. The catalogue lives in node config - versioned, reviewable, auditable. No prompt-injected tool names execute.

Allow-listedVersionedReviewable

STRONGLY_SERVICES routing

Every tool call resolves through STRONGLY_SERVICES. The agent never holds a database URL, an API key, or a service hostname directly. Rotation, RBAC, and audit happen at the platform layer.

No raw URLsCentralised authRotatable

Result formatter

The raw tool response (JSON, rows, blobs) goes through streaming-tool-result-formatter. The formatter renders human-readable prose for TTS - and a structured side-channel for downstream nodes.

Prose for TTSStructured side-channelEditable template

Streaming voice loop

STT, LLM, tool-router, tool-executor, formatter, and TTS all run behind the typed-frame contract. The user hears 'looking that up…' before the tool returns.

ConcurrentTyped framesADR-S11

Single-shot today, loop next

The current pattern is single-shot: one tool call per turn. Multi-step loops await runtime cycle support - same graph contract, just unblocked when ADR-S16 cycles ship.

Single-shotADR-S16-readyHonest scope

Live span tree

Tool name, argument shape, latency, and result size land on every span. Find slow tools, broken endpoints, or arg drift in the canvas overlay.

ADR-S14Per-toolArgument trace

Built on

Real services. Your stack.

Every dependency is a registered Strongly service or a model you control. Swap any one of them in the install wizard. The graph stays intact.

Tool router

streaming-tool-router - catalogue-validated

Tool executor

streaming-tool-executor - STRONGLY_SERVICES-routed

Formatter

streaming-tool-result-formatter - editable templates

LLM model

Function-calling chat model (gpt-4o default)

Five common customisations

Tune it. Don't fork it.

The marketplace template is the graph. Every customisation below is a config change or a single-node addition - never a rewrite.

Add a tool

Drop a new datasource into the workflow's STRONGLY_SERVICES, register it in the tool catalogue, redeploy. The LLM picks it up on the next session.

Restrict callable tools

Per-org or per-role allow-lists in the tool-router config. The same workflow can serve admins (full catalogue) and end-users (read-only subset).

Custom result format

Edit the tool-result-formatter template per tool. Tabular results → 'top three ranked by X'. Single-row results → narrative. Errors → graceful apology.

Caching

Insert streaming-tool-cache between the router and the executor. Repeated calls within the TTL skip the network hop entirely.

Audit-first deploy

Add streaming-conversation-store tagged with tool name and result hash. Compliance teams replay every tool invocation.