Open standards and tooling for AI-assisted engineering.
The dispatch pipeline that makes AI coding agents reliable, measurable, and cheap to run at scale. Composable open-source components — pick the ones you need.
Most AI coding tools optimize a single agent doing a single task. That's the wrong unit of work.
The real bottleneck is orchestration: dispatching dozens of agents in parallel, giving each just the context it needs, verifying their output before merge, and learning from every dispatch to make the next one better.
standra.ai is the open layer for that. Each phase of the pipeline has a focused tool. Use them together or pick one.
| Phase | What happens | Component | Repo |
|---|---|---|---|
| Task | Structured I/O — typed fields replace free-form prose. ~200 tokens vs ~2000. | CACP | zenprocess/cacp |
| Context | 5-tier submodular file assignment, task-aware retrieval, cache-aware optimization. | Sieeve | open-source release pending |
| Context (rules) | Compact, tabular, LLM-native rule definition language. The format Sieeve compiles into. | Axiom | zenprocess/axiom |
| Dispatch | Tracker-agnostic, git-native multi-agent dispatcher. Pre-dispatch doctor, response gate, AC compliance, regression detection. | (internal — release pending) | personal R&D |
| Verify | 4-dimensional benchmark harness — multi-turn, multi-agent, parallel dispatch with tool calling. | PawBench | zenprocess/pawbench |
| Verify (serving) | Model registry for optimized LLM serving configurations. The "spec sheet" for self-hosted vLLM. | ServingCard | zenprocess/servingcard |
| Learn | Mines Claude Code session history for friction, patterns, gaps. Evolves your skills via autonomous experiment loops. | Afterburn | zenprocess/afterburn |
A single dispatch, end-to-end:
- Task comes in (GitHub, Linear, GitLab — tracker-agnostic), wrapped in CACP
- Sieeve compiles task-aware context (Axiom rules + 5-tier file assignment)
- Pre-dispatch doctor rejects bad dispatches instantly (zero cost)
- Agent executes in isolated git worktree (Claude or local Hermes/vLLM)
- Gate stack runs: response gate (9 structural checks), test gate, AC compliance, regression detection, verify
- GEPA (Goal/Effort/Plan/Assumptions) scores the trajectory — 4-dimensional reflective optimization
- On pass → merge. On block → retry/partial. On doctor fail → rejected
- Telemetry is mined by Afterburn to improve tomorrow's dispatch
Every layer is measured.
| Layer | Tool | Savings |
|---|---|---|
| System prompt | Axiom compiled.ctx | 5000 → 500 tokens |
| Protocol | CACP structured fields | 2000 → 200 tokens/response |
| File context | Sieeve 5-tier submodular | Task-aware: FULL/SKELETON/SUMMARY/MANIFEST/OMITTED |
| Single file | Manifest vs inline | 50–100x per file |
| Mid-trajectory | AgentDiet pruning | 40–60% context reduction |
Validated: 87% token savings at parity on 684 compliance runs.
Most AI coding work today is single-shot, manual, and unmeasured. Token costs are opaque. Quality is self-reported. There's no feedback loop.
standra.ai makes the whole pipeline:
- Measurable — every dispatch produces structured telemetry
- Cheap — context compilation is the difference between $0.50 and $5.00 per dispatch
- Reliable — gates catch false-positive "done" claims before they hit your branch
- Self-improving — Afterburn evolves your skills based on what actually worked
standra.ai stands on the shoulders of researchers and builders who shipped the foundational ideas:
- GEPA — Goal/Effort/Plan/Assumptions reflective prompt optimization. The dispatcher scores every dispatch on these 4 dimensions and feeds the verdict back into routing. Agrawal et al., 2026
- RLM (Recursive Language Models) — Zhang, Kraska & Khattab. The REPL-driven recursive analysis pattern that lets Afterburn process 90MB+ session files no context window can hold.
- CCAR (Claude Code AutoResearch) — Mitko Vasilev. The autonomous experiment loop Afterburn uses to evolve skills (mutate → benchmark → keep/discard → commit).
- Andrej Karpathy — Software 2.0, LLM OS, and the public thinking on agentic systems that frames why orchestration is the real engineering problem, not the model.
- addyosmani/agent-skills — Addy Osmani. The Save Point Pattern, anti-rationalization tables, and skills-as-gated-workflows pattern that shape how the dispatcher structures dispatch prompts.
- DSPy — Stanford NLP. The original GEPA optimizer and the systematic approach to LM programming that influenced CACP's structured I/O design.
- LLMLingua — Microsoft Research. Prompt compression research underpinning Sieeve's tiered context strategy.
If we missed your work and you should be on this list — open an issue.
- ⭐ Star the components that interest you
- 🐛 File issues — every open-source project welcomes them
- 💬 Discussions are open
- 📧 hello@standra.ai
Always happy to swap notes with people working on similar problems → cal.com/vvladescu (15min AI/architecture chat).
standra.ai — Open Standards for the AI-Ready Enterprise