Nine agents, organized like a real R&D team, build a DSP chip toolchain
An embedded-software developer set out to build a complete programming toolchain for the TI C66x DSP — IDE, compiler, assembler, linker and automated tests — the kind of multi-module, multi-repo systems project that normally takes a whole team. He organized 9 Codex agents as that team; over 20 days they processed 536 events.
Background
One person building a full toolchain hits four walls at once: parallel development across interdependent modules; heavy test-and-debug cycles (TI DSS tests, regression runs) that demand deep knowledge of each module; sequencing work across dependencies (compiler output feeds the assembler feeds the linker); and the constant risk of editing the wrong repo when many modules change together.
The bottlenecks
Parallel multi-module pressure
Compiler, assembler, linker, IDE and test tooling are independent modules that must work together — brutal context-switching for one person.
Testing and debugging load
Each bug requires deep knowledge of one module; nobody can hold all modules in their head at once.
Coordination across dependencies
Modules depend on each other, so the order and rhythm of development and testing needs careful arrangement.
Cross-repo mistakes
Synchronized changes across repos invite wrong-repo edits and version drift.
How it unfolded
Phase 1 — Build the AI team, split the work (days 1–5)
One master agent handles system integration and roll-up; eight worker agents own the compiler, assembler, linker, IDE and test automation. Each agent is bound to its own repository and working directory with clear responsibility boundaries, so changes never collide. The master decomposes the project goal into subtasks and assigns them.
Phase 2 — The test-and-fix loop (days 6–12)
Once first versions landed, the test agent ran TI DSS and regression suites; the master routed each bug to the agent owning that module and tracked the fix. The loop — test finds, master assigns, worker fixes, test re-verifies — ran like a miniature R&D team's cadence.
Phase 3 — Integration, and where it strained (days 13–20)
The master integrated compiler, assembler and linker outputs and drove toward passing TI DSS end-to-end. But at scale the seams showed: Windows permission errors blocked file access, some workers lost contact with the master, task hand-offs failed so bugs went unassigned, cross-agent edits landed in the wrong repo, and concurrent scripts conflicted. The user eventually switched to one agent per project, and usage dropped — the stability issues were likely why large multi-agent collaboration didn't continue.
Inside the Workspace
What changed
Team-structured AI division of labor works
Nine agents were successfully organized like a real team — separate repos, modules and test duties — running the full assign → develop → test → fix → integrate workflow.
High per-module velocity
With clear ownership, each agent developed and debugged its module quickly.
The master as project manager
The master assigned tasks, tracked progress and rolled up results; the user only set goals and arbitrated big calls — no per-module micromanagement.
Takeaway
This was a hard engineering experiment: a virtual R&D team of 9 agents building a DSP toolchain. Its biggest value is the proof that developers can mirror real organizational structure — repos, modules, test ownership — onto agents, offering a reference model for "one person + many agents" on mid-sized-and-up projects, along with an honest picture of where multi-agent stability still matters.
Put agents to work in your team
Start from one painful, repetitive workflow — the way the teams in these stories did.
