Agentic coding with Areg: try it, break it, send me the run

[aregtech]

Can a coding agent that has never seen a framework build a correct, multi-process C++ application from that framework's own documentation, in one session?

With Areg, that is a design goal, and I'd like you to put it to the test with your own agent, your own model and your own machine.

Why Areg and coding agents fit together

Most frameworks are documented for humans. An agent working with them leans on training data that is often years out of date, then guesses, compiles, fails and tries again. Areg is built so an agent learns it from the repository, during the task:

• 🤖 AI-friendly. The agent describes what the system does, not how bytes move. It writes a short design (services, data types, a state machine); generators turn it into .siml, .dtml and .fsml documents and C++ skeletons; the agent writes only the method bodies. Threads, IPC, service discovery and reconnection are the framework's job, not the agent's.
• 🧭 AI-ready. AGENTS.md routes each task to one page. A contract checker catches a renamed or invented API name before the compiler does, and a scenario runner proves the programs work together end to end. CI checks the documentation against the code, so the pages an agent reads cannot quietly drift away from the framework.
• 🪙 Token-optimized. Every page of the agent documentation has a size budget that CI enforces, and the whole reading set has a hard ceiling. The agent reads the pages it is routed to, not the source tree.

This appropriate improvements are done since HEAD 650cd619577180a109cb2b85046a8c3fd93ecbe1 and the full support is planned since version 2.0.0.

What that gives in our own runs (method and ranges):

• ✅ 57 of 57 runs passed every hidden probe over 3 days, on 4 tasks and 3 model families.
• ⚡ With Claude Code and Sonnet 5, about half the API requests of gRPC (at least 30% less), for equal output tokens and time, against a framework the model was trained on.
• 🎯 Zero filesystem searches. The documentation answered what the agent needed.

See what an agent produces

Two examples in the repository were built entirely by an AI agent and are kept as references:

Example	What it is
33_tempalarm	a temperature monitor publishes readings and raises a threshold alarm; an operator process subscribes and reacts
34_coffeemachine	a coffee machine driven by a generated state machine, with a simulated user in a second process

Read the code and judge it yourself. The .fsml state machine opens in Lusan, which draws it, so you can check the logic against what was asked.

Try it yourself

examples/ai-benchmark/ holds 8 ready task prompts (temperature alarm, coffee machine, ATM written by hand and as a state machine, print/scan device, elevator, sensor gateway, greenhouse) and a harness that:

1. runs your agent cold, in an isolated snapshot of the SDK, with the web off;
2. lets it design, generate, build and test the application;
3. scores the result with hidden probes the agent never sees: wrong start order, a peer that never comes, a peer that dies mid-run, busy-waiting.

Clone master (the agent documentation is newer than the last release), then start from an empty directory outside the checkout:

mkdir -p ~/runs && cd ~/runs
/path/to/areg-sdk/examples/ai-benchmark/run-benchmark.sh --task examples/ai-benchmark/prompt-tempalarm.md --attempts 15
/path/to/areg-sdk/examples/ai-benchmark/run-benchmark.sh --task examples/ai-benchmark/prompt-coffeemachine.md --attempts 15

mkdir C:\runs; cd C:\runs
C:\src\areg-sdk\examples\ai-benchmark\run-benchmark.ps1 --task examples/ai-benchmark/prompt-coffeemachine.md --attempts 15

• Agents: Claude Code is the default. --agent copilot, --agent codex and --agent gemini select the other CLIs, --model and --effort the rest. Cursor or any IDE agent can run the same prompt by hand.
• You need: a C++17 compiler, CMake, Python 3, and one agent CLI you are logged in to.
• Scale: our first measured coffee-machine run took about 15 minutes and cost about $2 with Claude Code and Sonnet 5. One run; yours will differ.
• Curious how it compares? Add --framework grpc and the same task runs on gRPC.

Share your run 🙏

This is the part I care about most. A failed run is worth more to me than a passed one: every failure points at a page, a generator or a framework path I can fix.

Each run leaves a directory such as 20260923c-coffeemachine/. Leave out the SDK snapshot and the build output, and what remains is about 30 KB:

tar -czf 20260923c-coffeemachine.tgz --exclude=20260923c-coffeemachine/sdk --exclude=20260923c-coffeemachine/work/build 20260923c-coffeemachine

Run it from the directory that holds the run, with your run's name. tar also ships with Windows 10 and later, so the same line works in PowerShell. What is inside, and why it helps:

File	Tells me
meta.txt	agent, model, effort and the SDK commit it read
prompt.txt	exactly what the agent received
run.out, run.err, result.json	the agent's own report and the usage record of its CLI
verify.json	each hidden probe's verdict, with its evidence
work/	the project the agent wrote
sdk-head.txt, sdk-md5.txt	which SDK tree it read, so sdk/ itself is not needed

Attach the archive to a reply here, with a few lines:

Agent / model / effort :
Task                   :
Hidden probes          :   of 5
What surprised you     :
Where the agent guessed:

prompt.txt and meta.txt contain your local paths, including your user name. Have a look before you post. If a run cannot be shared publicly, send it to info[at]areg[dot]tech.

Ran it by hand, without the harness? The prompt, the agent's final report and the project directory without its build output are just as useful.

What happens with it

We go through every shared run and turn each finding into a fix: a page an agent misread, a generator that wrote the wrong line, or a framework path that behaves badly under a probe. I'll reply in your thread with what was found and link the fix when it lands.

Also welcome

• A task prompt of your own. The prompts name no framework, so the same task can run on Areg and on gRPC. How to write one.
• A model or agent we have not tried. Every new combination tells me something.
• Your opinion. What would it take for you to trust AI-written C++ in a multi-process system?

If Areg saved your agent API requests, a ⭐ on the repository helps the next developer find it.