Frequently Asked Questions About Saraev AI Agent Orchestration System
22 answers covering everything from basics to advanced usage.
// Basics
What does 'agentic AI' actually mean in practice?
Agentic AI refers to systems that autonomously run a continuous Observe-Think-Act loop toward a goal, rather than answering a single prompt. An agent reads context, reasons about the objective, calls tools or edits files, then repeats — accumulating context each cycle until a Definition of Done is satisfied. This lets it complete multi-step tasks like building an app without step-by-step human prompting.
What's the difference between an orchestrator and a sub-agent?
An orchestrator is the top-level agent that receives the goal, plans, delegates subtasks, and validates recombined outputs. Sub-agents are spawned by the orchestrator to work on a specific subtask in their own isolated context windows, reporting results back for aggregation or verification. Claude Code is the default orchestrator; Gemini and Codex often serve as specialized sub-agents.
What is MCP and why does it matter for orchestration?
MCP (Model Context Protocol) is the communication standard used to register and query different AI models and tools as servers within a multi-agent system. It matters because it lets a single orchestrator route subtasks to Claude, Gemini, or Codex through a common interface, enabling true cross-provider orchestration rather than juggling separate tools manually.
What is 'traversing the search space' and why does it matter?
Traversing the search space is the mathematical idea behind stochastic consensus: running multiple parallel queries covers more of the total space of possible answers than a single query. It matters because valuable, rare solutions the model would seldom produce on any single run surface when you deliberately spread many varied agents across the space at once.
// How To
How do I write a good Definition of Done?
Write specific, measurable completion conditions in plain language. Instead of 'build the app,' use 'app renders in browser with no console errors, API responds to requests, and image generates on demand.' For research, use '10+ empirical sources compiled with citations.' Place it as the final section of every prompt so the agent knows exactly when to stop looping.
How do I set up global versus local instruction files?
Maintain two layers: a global agents.md holding user-wide preferences applied across all projects, and a local project-specific .md for project-only rules. Stack skills and inline prompts beneath them. This collapses large amounts of context into few tokens, preserving model quality. The global file carries persistent preferences; the local file carries project constraints.
How do I run stochastic multi-agent consensus correctly?
Take your core question and write N versions (3–10) with different analytical framings — conservative budget, measurable-only, end-user perspective, contrarian. Spawn all N simultaneously as sub-agents, each in its own context window. The orchestrator collects responses and calculates mode (consensus), divergence, and outliers, then synthesizes a consensus map. Critically, run them in parallel, never sequentially.
How do I trigger my instruction file to update at session end?
Before ending the session, prompt the agent: 'Update the Learned Rules section of [gemini.md/claude.md/agents.md] with any corrections, preferences, or patterns identified in this session.' The agent appends new rules so the next session starts with accumulated knowledge. This is the mechanism that makes error rates decline over time.
Can I build my own custom skill file, and what should it contain?
Yes. A skill file needs a title section with three-hyphen delimiters, a name, a description, optional tools, and a step-by-step procedure. Store it in your workspace skills folder and invoke it by name in prompts. Keep it concise — overly long skills degrade model quality and raise cost. Scope each skill to one repeatable workflow.
// Troubleshooting
My agent stops halfway through the task — how do I fix it?
This usually means your Definition of Done is missing or ambiguous. Add explicit, measurable completion conditions to the prompt so the agent knows it hasn't finished. If it stops arbitrarily despite a clear Definition of Done, check whether context has degraded from over-accumulation — collapse rules into skills and use sub-agents with isolated context windows.
Why do the outlier ideas from consensus runs sometimes seem wrong?
Outlier wild cards appear in only 5–10% of stochastic consensus runs and may be hallucinations rather than breakthroughs. Never treat them as automatically valid. Evaluate each outlier individually against evidence before acting, while acting more confidently on consensus items that appear in the majority of agents. Divergent items require careful reasoning before adoption.
My instruction file has hundreds of rules and quality is dropping — what now?
Rules begin to conflict when a single file becomes unwieldy (hypothetically 1,000+ entries), and longer context degrades model quality while raising cost. Prune redundant or outdated rules, restructure by category, and move project-specific rules into local .md files. Collapse repeatable procedures into concise skills rather than inflating the instruction file.
Why is my multi-agent API bill much higher than expected?
Subsidized platform plans (like Claude Max at $200/month) do not cover multi-model API orchestration — API usage is billed separately per provider and can cost significantly more. Each loop iteration also stacks more tokens into context, compounding cost. Budget for per-provider API keys, and avoid multi-agent orchestration until simple single-agent workflows work first.
// Comparisons
How does this system compare to a generic 'just prompt ChatGPT' approach?
Generic prompting relies on one model, one context window, and manual re-querying. This system parallelizes across specialized models, exploits stochasticity through consensus, and self-corrects via instruction files that persist across sessions. The result is wider solution-space coverage, more consistent outputs, and declining error rates — at the cost of more setup and higher token spend.
How does stochastic consensus compare to just re-running the same prompt?
Sequential re-querying takes N times as long and loses the core time advantage, while covering less of the solution space than deliberately varied, parallel framings. Stochastic consensus spawns N sub-agents simultaneously with different analytical angles, traversing more of the total answer space at once and surfacing rare wild cards that repeated identical queries rarely produce.
Should I use Claude Code or Anti-Gravity for a frontend build?
Use Gemini/Anti-Gravity for frontend and design output — it produces superior UI and has fast, multimodal strengths. Claude Code is weaker at frontend design but stronger at orchestration and interpretability. In a multi-model build, use Claude Code as orchestrator and delegate the frontend subtask to Gemini via MCP for the best of both.
How does a skills file differ from an instruction file?
A skills file encodes a specific repeatable workflow (like video-to-action) invoked by name only when needed, forcing a deterministic procedure. An instruction file (agents.md) is prepended to every session and holds persistent preferences and learned rules. Skills reduce variance for recurring tasks; instruction files carry cross-session memory and alignment.
// Advanced
How do sub-agent verification loops improve output quality?
After sub-agents complete work, route their outputs to a separate agent or the orchestrator for real-time peer review against the original Definition of Done. A reviewing agent catches errors, inconsistencies, and missed requirements that a single agent's self-review misses. If issues surface, loop back to the responsible agent — improving quality without needing human review of every step.
How do I keep context windows clean in long agent runs?
Practice context window hygiene: collapse knowledge into concise skills and instruction files, use sub-agents with their own isolated context windows so subtask tokens don't pollute the orchestrator, and avoid over-accumulating rules in one file. Since context grows every loop iteration — raising cost and eventually degrading quality — proactive pruning preserves both performance and budget.
What are agent chat rooms and when should I use them?
Agent chat rooms are shared contexts where multiple agents communicate, debate, and build on each other's work, pushing collective output beyond any single agent's. Use them for coordinated parallel tasks — like a fleet of agents contacting separate leads that must avoid duplication, or ideation where agents refine each other's proposals in real time.
When is multi-agent orchestration overkill?
It's overkill for most everyday tasks. Since model differences are small percentage points and orchestration adds token cost, latency, and complexity, only orchestrate when bleeding-edge quality on distinct subtask types is critical — like a full-stack build needing top-tier frontend, backend, and testing simultaneously. Get single-agent workflows working reliably first.
How do I use Claude's interpretability to steer an agent mid-run?
Claude Code exposes a reasoning tab showing the Think step in real time. Watch it to verify the agent is on the right path, catch wrong assumptions early, and inject new resources or constraints without restarting. Gemini and Codex are less interpretable — treat them as fire-and-forget, setting the target and Definition of Done precisely before launch.