Overview
The model is one layer. The differentiation is the orchestration around it. Agent Orchestration is the runtime that turns Embedder from a chat surface into a verification platform: ingesting evidence, planning a specification before writing code, executing inside a controlled toolchain, and validating against real hardware in a closed loop — with subagents, audit trails, and human gates where they belong.
What this white paper covers
- The four layers — evidence ingestion, hardware intelligence, toolchain execution, hardware verification.
- The closed loop — build, flash, test, debug as one cycle.
- Specification-first mode — plan with datasheet citations before code.
- Subagents — parallel work under one orchestrator.
- Human-in-the-loop — approval policy you control.
- Audit and replay.
The four layers
- Evidence ingestion — docs, schematics, errata, SVD device files, build output, traces, registers, board context.
- Proprietary firmware intelligence — indexed hardware knowledge, component mappings, constraints, and retrieval logic across 500+ MCUs and 3,000+ peripherals.
- Toolchain execution — generate, build, lint, flash, test, iterate inside controlled engineering flows.
- Hardware verification loop — validate assumptions against documentation, target behavior, and real-silicon results.
The closed loop
Build → flash → test → debug, run as one continuous cycle. The agent doesn't hand you a suggestion and walk away; it flashes the board, observes the result through the integrated serial terminal, ingests GDB state, and refines the code or the plan based on what actually happened. When the observed behavior diverges from the documentation, that divergence becomes the next thing to debug.
Specification-first mode
For non-trivial work the orchestrator runs a planning phase first: research the relevant docs, draft a structured implementation spec with citations from the reference manual, and surface assumptions for review. You review the plan — not just the diff. This is how teams catch agent mistakes before the code is written, not after. It's also how compliance teams get traceability: every register choice, init sequence, and timing decision links back to a specific paragraph in the source documentation.
Subagents
Real firmware work isn't one task. A bring-up has a driver to write, a HIL test to author, and a power profile to validate — partly in parallel. Subagents run those tracks concurrently under one orchestrator, with shared evidence, isolated tool access (one agent gets JTAG, others wait), and a unified audit trail. You see the whole tree of work in one place.
Human-in-the-loop
Most teams want autonomy on the boring work and approval on everything else. Orchestration lets you set that policy explicitly: which agents can land code without review, which need a human to approve a diff, which need explicit sign-off before touching flash on a customer board. Approvals show up where your team already works — PR comments, Slack, or the Embedder console — and the agent waits.
Audit and replay
Every prompt, tool call, file read, board interaction, and human approval is logged. The agent's reasoning is captured with its citations back to source documentation. Six months from now, when a regulator or a customer asks why a particular register value ended up in shipping firmware, you can trace it end to end. Sessions can be re-played against a fresh checkout for debugging.
Who it's for
- Teams running multiple agents at once — the cost of chaos compounds fast without coordination.
- Regulated programs — ISO 26262, IEC 62304, MISRA-C, DO-178C workflows where audit trails are mandatory.
- Enterprise hardware teams running Embedder across multiple product lines under one platform team.
Getting started
Orchestration ships as part of Embedder by default. For a walkthrough of the planning, subagent, and approval policies against your existing repo, talk to an engineer.