Fuze SDK · Agent · @fuze-ai/agent

A TypeScript agent framework with the compliance layer wired in.

Same primitives as Guard — guard, loop, budget, side_effect, approval — but as a build-your-own runtime. Build a new agent on Fuze Agent and the Fuze Control dashboard starts receiving evidence on the first run. Inspired by Mastra and just-bash; aimed at small EU teams writing agents from scratch.

Framework docs View on GitHubnpm i @fuze-ai/agent

Primitives

The runtime is the compliance layer.

Each primitive is a function you compose into your agent. The evidence emits as you call them — there is no separate instrumentation step.

tool()

Art. 12
@fuze-ai/agent

Typed tool definitions with input/output schemas. Every call lands on the evidence stream with its arguments and return value.

tool: lookup
input  { id: string }
output  Record | null
schema  valid

loop()

runtime
@fuze-ai/agent

Detects when the agent re-walks the same call graph and pauses before the budget runs out.

cycle · paused

memory()

Art. 19
@fuze-ai/agent-memory

Append-only conversation and step memory wired into the same evidence chain as everything else.

step.01tool.lookup
step.02model.call
step.03tool.lookup
step.04model.call

side_effect()

runtime
@fuze-ai/agent

Marks calls that touch the outside world. The wrapper records the effect, the response, and any rollback hook you attach.

http.POSTlogged
db.writelogged
email.sendlogged

approval()

Art. 14
@fuze-ai/agent-durable

ctx.requestOversight() suspends the run on a durable adapter. The reviewer's signature binds the human decision into the chain.

queue · 1 pending
extend offer · candidate_4f1c

budget()

runtime
@fuze-ai/agent

Hard caps on tokens, steps, and wall-clock per run, with the breach event recorded on the evidence stream.

tokens7,238 / 10,000
steps9 / 20

A first agent

Two imports, one run.

Build an agent with a tool, a model, and a budget. The Control dashboard receives the run's evidence as soon as it finishes. EU residency providers (Mistral, Scaleway, OVHcloud) ship in @fuze-ai/agent-providers.

  • Typed tools with Zod schemas
  • EU residency providers in @fuze-ai/agent-providers
  • Durable approvals via @fuze-ai/agent-durable
  • Evidence emits to a local store or your Fuze Control tenant
agent.ts
import { z } from 'zod'
import { quickAgent, quickTool } from '@fuze-ai/agent/quickstart'
import { mistralModel } from '@fuze-ai/agent-providers'

const lookup = quickTool({
  name: 'lookup',
  input: z.object({ id: z.string() }),
  run: async ({ id }) => fetchRecord(id),
})

const screener = quickAgent({
  model:  mistralModel('mistral-large-2'),
  tools:  [lookup],
  budget: { tokens: 10_000, steps: 20 },
})

const result = await screener.run(prompt)
signed · Art. 12, 14, 26 ← evidence chain

Article coverage

What the framework gets you out of the box.

Fuze Agent inherits the SDK’s evidence model. The split below mirrors the public fuze-ai README — same shipping/partial/roadmap as the SDK page.

Shipping

Art. 12Automatic event recordingEvery tool call, model call, and decision lands as structured JSONL through the framework runtime.
Art. 14Human oversight`ctx.requestOversight()` suspends the run on a durable adapter; reviewer signature binds the human decision into the chain.
Art. 26Deployer monitoringRoutes to Fuze Control for run feeds, agent health, retention, and admin audit.
Art. 27FRIA`@fuze-ai/agent-fria` compiles a FRIA draft from evidence span attributes.
Art. 73Serious incident reporting`@fuze-ai/agent-incident` drafts a regulator-ready Article 73 report from the same stream.

Partial today

Art. 19Log retentionAppend-only event store with configurable retention. HMAC-SHA256 hash chain is in the Python SDK; TypeScript parity is on the roadmap.
annex ivTechnical documentation`@fuze-ai/agent-annex-iv` maps evidence span attributes to Annex IV sections and ISO 42001 controls.

On the roadmap

Art. 9Risk-management systemTooling for residual-risk logs and control-registry tracking.
Art. 10Data & data governanceSource lineage and drift logging for inference and training data.
Art. 13Transparency to deployersCapability cards and a known-limitation registry exported with each run.
Art. 15Accuracy & robustnessEval-suite integration and jailbreak-attempt logging on the hot path.
Art. 72Post-market monitoringDrift alerts and performance-regression diffs per deployment.

Open source

Same MIT licence as the SDK.

Fuze Agent ships in the same monorepo as the rest of the ecosystem. Run it self-hosted with a local evidence store, or point it at the managed Fuze Control tenant.

Next

Five minutes to know if the Act applies to you.

The classifier walks you through Annex III and tells you which Articles apply, which tier of risk you sit in, and what evidence the Fuze Control dashboard would start emitting on day one. Free, no signup, no email gate.