AX Agent Orchestrator: 6 Production Workflows Founders and Operators Can Build Now

AX, Google’s open agentic orchestrator, surfaced as a high-relevance workflow item on September 21, 2026 because the agent market has crossed a practical line: teams are no longer asking whether a model can answer a prompt. They are asking how to turn messy, multi-step work into repeatable pipelines with planning, tool calls, verification, retry logic, and model routing.

That matters for founders, operators, and developers because the expensive part of agent deployment is not the first demo. It is making the workflow traceable enough to trust, cheap enough to run daily, and flexible enough to swap models as requirements change. AX fits the current market need because it gives teams an orchestration layer: planner stages, tool-use stages, evaluator stages, fallback-model stages, and structured execution traces that can be inspected when something goes wrong.

This guide focuses on what you can do next. We will map AX-style orchestration patterns to six practical workflows: research operations, support triage, data enrichment, coding review, back-office automation, and executive reporting. We will also show two copyable step-by-step pipelines, compare model choices, and estimate what each run costs using current model prices from AI Cost Check.

---

What changed: agents are becoming production workflows

The most important change is architectural. The first wave of AI apps wrapped a single model call in a chat interface. The current wave uses agents to perform work across documents, APIs, tickets, repos, CRMs, spreadsheets, browsers, and internal tools. That creates three new requirements:

1. Traceability: Every decision needs a record: which model ran, what tools were called, what evidence was used, and why an output passed or failed.
2. Control: Operators need guardrails, retry policies, structured outputs, and human approval gates.
3. Cost routing: Not every step deserves a premium model. Extraction, classification, deduplication, and formatting can usually run on cheaper models.

AX is timely because it gives builders a way to define these stages explicitly. Instead of shipping a single giant prompt that asks an LLM to research, reason, browse, format, validate, and decide, you split the job into smaller stages:

• A planner decomposes the task.
• A retriever or tool-use stage fetches data.
• A worker model performs extraction, summarization, generation, or transformation.
• An evaluator checks correctness, confidence, schema compliance, policy fit, or business rules.
• A fallback stage escalates hard cases to a stronger model or routes low-confidence items to humans.

This is exactly where agent costs become manageable. A premium model such as GPT-5.2 pro costs $21 per 1M input tokens and $168 per 1M output tokens. A cheap fallback or bulk-processing model such as Gemini 2.5 Flash-Lite costs $0.10 per 1M input tokens and $0.40 per 1M output tokens. The difference is not marginal; it determines whether you can run a workflow on 50 items per day or 50,000.

---

What AX makes possible: six workflows to build now

AX is best understood as a workflow skeleton. The orchestrator coordinates model calls and tools, but your business value comes from choosing the right sequence of stages. These six patterns are practical enough to copy into a startup, ops team, or internal platform.

1. Research operations pipeline

Use AX to convert open-ended research into a traceable evidence workflow. A research agent can gather sources, extract claims, score relevance, summarize findings, and produce a final memo with citations.

A good production version has five stages:

Stage	Purpose	Recommended model
Planner	Break the question into sub-questions	Gemini 3 Pro or GPT-5.2
Retrieval	Search internal docs, web, or database	Tool/API stage
Extraction	Pull facts, dates, companies, numbers	Gemini 2.5 Flash
Evaluator	Check source quality and contradictions	Claude Sonnet 5
Final memo	Produce executive-ready output	GPT-5.2

The orchestration advantage is that every final claim can be traced back to retrieved evidence. That makes this workflow useful for investor research, market maps, competitive analysis, vendor diligence, and product strategy.

2. Support triage and resolution drafts

Support workflows are perfect for orchestration because they involve classification, policy lookup, user history, knowledge-base retrieval, escalation rules, and draft responses. A single prompt can handle a toy ticket. A production ticket queue needs a pipeline.

AX can route tickets through:

• Intent classification
• Severity scoring
• Account lookup
• Policy retrieval
• Suggested response generation
• Safety and tone evaluation
• Escalation if confidence is low

For cost control, classify and retrieve with a cheaper model such as GPT-5 nano at $0.05/$0.40 per 1M tokens or Command R at $0.15/$0.60 per 1M tokens. Save stronger models for escalations and sensitive customer communications.

3. Data enrichment for sales and growth teams

Messy CRM data is one of the highest-ROI places to use an agent orchestrator. AX can take a raw company record, enrich it with website data, classify the segment, find relevant contacts, summarize buying triggers, and output a structured row.

A practical enrichment flow:

1. Validate company domain.
2. Fetch website, LinkedIn-like metadata, or internal account history.
3. Extract company size, category, ICP fit, region, pricing page signals, and technology stack hints.
4. Score fit against your sales criteria.
5. Generate a short outreach angle.
6. Flag uncertain fields for review.

The key is schema enforcement. Enrichment workflows should never return freeform prose as the primary output. Use JSON fields such as company_category, icp_score, evidence_urls, confidence, and human_review_required.

4. Coding review and pull request analysis

Coding agents are moving from “generate this function” to “review this pull request, run checks, identify risky changes, and propose minimal fixes.” AX-style orchestration helps because code review requires multiple passes:

• Diff summarization
• Test impact analysis
• Security scan
• Style and maintainability review
• Suggested patch generation
• Final evaluator pass

Use Codex Mini at $1.50 input / $6 output per 1M tokens or GPT-5.3 Codex at $1.75/$14 per 1M tokens for code-heavy stages. Escalate only high-risk architectural or security reviews to GPT-5.6 Cyber, which costs $12.50/$75 per 1M tokens.

5. Back-office document automation

Finance, legal ops, HR, and procurement teams run repeatable document workflows: invoices, contracts, vendor forms, compliance packets, onboarding documents, and internal approvals. AX can turn these into auditable pipelines.

A document agent should not just “read and answer.” It should:

• Classify document type.
• Extract required fields.
• Compare terms against policy.
• Detect missing information.
• Produce an approval recommendation.
• Route exceptions to humans.

This is where an evaluator stage matters. If an invoice is missing a PO number, a vendor contract has non-standard payment terms, or an HR document lacks required signatures, the pipeline should stop and create a review task.

6. Executive reporting and weekly business reviews

Founders and operators spend hours merging metrics, ticket themes, sales notes, product updates, and financial signals into weekly updates. AX can coordinate the data pulls and narrative generation.

A strong reporting workflow includes:

• Data connector stages for CRM, analytics, support, and finance tools
• Metric validation against prior periods
• Anomaly detection
• Narrative generation
• Risk and action-item extraction
• Final executive summary

Use cheaper models for metric labeling and summarization, then a premium model for the final narrative. The final document often goes to leadership, investors, or customers, so the polish and reasoning quality justify a stronger model.

---

Copyable workflow 1: research operations agent

This workflow turns a messy research question into a sourced memo. It is designed for founders, analysts, and product teams that need reliable answers instead of a pile of links.

Goal

Input: “Should we build an integration with Tool X for mid-market SaaS companies?”

Output: A structured memo with market evidence, customer demand signals, competitor coverage, implementation complexity, risks, and recommendation.

Step-by-step pipeline

Step	AX stage	What happens	Model/tool
1	Intake parser	Convert the question into decision criteria	GPT-5.2
2	Planner	Generate 5-8 research sub-questions	Gemini 3 Pro
3	Retrieval	Search web, internal docs, CRM notes, tickets	Tools/APIs
4	Evidence extractor	Pull claims, metrics, quotes, source URLs	Gemini 2.5 Flash
5	Contradiction checker	Identify conflicting evidence	Claude Sonnet 5
6	Decision scorer	Score demand, strategic fit, complexity, risk	GPT-5.2
7	Memo writer	Produce final recommendation with citations	GPT-5.2
8	Evaluator	Check missing citations and unsupported claims	Claude Sonnet 5

Suggested prompt structure

For the planner stage:

You are planning a business research workflow.
Decision: {{decision_question}}
Company context: {{company_context}}

Return JSON with:
- sub_questions: 5-8 questions
- required_sources: source types to retrieve
- decision_criteria: weighted criteria
- disqualifying_evidence: evidence that would change the recommendation

For the evaluator stage:

Evaluate this memo for unsupported claims.
For each claim, mark:
- supported: true/false
- evidence_id
- confidence: high/medium/low
- required_fix

Reject the memo if any recommendation claim lacks evidence.

Cost estimate

Assume one research run uses:

• Planner: 8,000 input tokens, 1,000 output tokens
• Retrieval/extraction across documents: 60,000 input, 8,000 output
• Evaluator passes: 20,000 input, 2,000 output
• Final memo: 15,000 input, 3,000 output

A balanced stack using GPT-5.2, Gemini 2.5 Flash, and Claude Sonnet 5 lands around $0.25-$0.60 per research memo, depending on retrieval volume and output length. Running 1,000 memos would cost roughly $250-$600 in model usage.

A cheaper stack using Gemini 2.5 Flash-Lite for extraction and GPT-5 mini for drafting can push many research tasks below $0.10 per memo, but use it for internal scans, not board-level decisions.

---

Copyable workflow 2: support triage agent

This workflow handles high-volume support queues where every ticket needs classification, policy context, suggested action, and escalation logic.

Goal

Input: raw customer ticket, account metadata, product usage, and knowledge-base articles.

Output: category, priority, confidence, recommended next action, draft response, and escalation status.

Step-by-step pipeline

Step	AX stage	What happens	Model/tool
1	Ticket classifier	Categorize issue and detect urgency	GPT-5 nano
2	Account lookup	Pull plan, ARR, SLA, recent incidents	CRM/support API
3	KB retriever	Fetch 3-5 relevant support articles	Search/vector DB
4	Resolution planner	Choose action path	Gemini 3 Flash
5	Draft response	Generate customer reply	Claude Sonnet 5
6	Policy evaluator	Check refunds, legal, security, tone	GPT-5.2
7	Escalation router	Send low-confidence or high-risk cases to humans	Rules + model score

Routing rules to copy

Use deterministic thresholds instead of asking the model to decide everything.

If severity = "security" or "data_loss", escalate.
If ARR > $50,000 and sentiment = "angry", escalate.
If confidence < 0.78, escalate.
If refund_amount > $250, require human approval.
If policy_evaluator = "fail", do not send response.

Cost estimate

A typical ticket triage run might use:

• Classification: 2,000 input, 200 output
• Retrieval summaries: 6,000 input, 800 output
• Draft response: 4,000 input, 700 output
• Evaluation: 3,000 input, 300 output

Using GPT-5 nano for classification, Gemini 3 Flash for planning, Claude Sonnet 5 for drafting, and GPT-5.2 for evaluation, the blended cost is usually around $0.02-$0.08 per ticket. At 10,000 tickets per month, that is roughly $200-$800 in model usage.

If you use a premium model for every stage, costs rise quickly. For example, GPT-5.2 pro at $21 input / $168 output per 1M tokens is excellent for high-stakes reasoning, but it is overkill for routine password resets, “where is my invoice,” plan-limit explanations, and basic bug routing.

---

Model Choice and Cost: premium where it matters, cheap where it scales

The main cost advantage of orchestration is selective model use. Single-model agents are simple, but they waste money by sending every token to the same model. AX-style workflows let you route each stage to the cheapest model that can do the job.

Recommended model stack by stage

Workflow stage	Best default	Cheaper fallback	Premium option
Classification	GPT-5 nano	Gemini 2.0 Flash-Lite	GPT-5.2
Extraction	Gemini 2.5 Flash	Gemini 2.5 Flash-Lite	Claude Sonnet 5
Long-context research	Gemini 3 Pro	Gemini 3 Flash	GPT-6 Astra
Coding review	Codex Mini	Codestral	GPT-5.3 Codex
Final executive writing	GPT-5.2	GPT-5 mini	GPT-5.2 pro
Policy evaluation	Claude Sonnet 5	Claude Haiku 4.5	Claude Fable 5.1

Current pricing snapshot

Model	Input / 1M tokens	Output / 1M tokens	Context
GPT-5.2 pro	$21	$168	1,000,000
GPT-5.2	$1.75	$14	1,000,000
Claude Sonnet 5	$2	$10	1,000,000
Gemini 3 Pro	$2	$12	2,000,000
Gemini 3 Flash	$0.50	$3	1,000,000
Gemini 2.5 Flash-Lite	$0.10	$0.40	1,000,000
GPT-5 nano	$0.05	$0.40	128,000
DeepSeek V4.1 Flash	$0.15	$0.60	1,000,000

When premium models are worth it

Use premium models for stages where a wrong answer is expensive:

• Final recommendation for an executive decision
• Legal, finance, compliance, or security interpretation
• Code review for authentication, payments, data access, or infrastructure
• Ambiguous customer escalations
• Multi-document synthesis with conflicting evidence
• Agent planning for high-value actions

When premium models are overkill

Use cheaper models for stages with narrow outputs:

• Intent classification
• Entity extraction
• JSON normalization
• Deduplication
• Document type detection
• Formatting
• First-pass summarization
• Batch enrichment fields

The practical rule: if the output can be checked with a schema, a regex, a lookup table, or an evaluator, start with a cheap model. If the output requires judgment, synthesis, or accountability, route to a stronger model.

[stat] 420x The output-token price gap between GPT-5.2 pro at $168/1M output tokens and Gemini 2.5 Flash-Lite at $0.40/1M output tokens

For more model-level comparisons, use AI Cost Check or compare common premium and budget choices like GPT-5 vs Gemini 3 Pro, GPT-5 vs DeepSeek V3.2, and Claude Opus 4.6 vs Gemini 3 Pro.

---

AX architecture pattern: planner, tools, evaluator, fallback

A reliable AX workflow should be designed as a state machine, not a chat transcript. Each stage has an input contract, output contract, model choice, retry policy, and logging behavior.

The production pattern

Component	What it does	Failure mode to handle
Planner	Breaks work into steps	Over-planning, irrelevant subtasks
Tool runner	Calls APIs, search, DB, browser, code tools	Missing data, rate limits, auth failures
Worker model	Performs extraction, writing, reasoning	Hallucination, schema drift
Evaluator	Scores output quality and risk	False confidence
Fallback router	Retries with stronger model or human	Cost spikes
Trace logger	Stores inputs, outputs, decisions, tool calls	Privacy and retention issues

The evaluator is the most underrated stage. Without it, the workflow has no built-in quality control. The evaluator should not be vague. Give it specific checks:

• Does the JSON match the schema?
• Are all claims supported by evidence IDs?
• Is confidence above threshold?
• Are restricted actions requested?
• Does this need human approval?
• Did the tool call return stale or incomplete data?
• Did the model cite data that was not retrieved?

Suggested evaluator output

{
  "pass": false,
  "confidence": 0.71,
  "failure_reasons": [
    "Recommendation lacks source evidence",
    "Customer refund policy not retrieved"
  ],
  "next_action": "retry_with_premium_model",
  "required_stage": "policy_retrieval"
}

A clean evaluator output lets the orchestrator make deterministic decisions. If pass=false and next_action=retry_with_premium_model, AX can escalate from Gemini Flash to Claude Sonnet 5 or GPT-5.2. If the failure is missing data, it can rerun retrieval instead of wasting money on a bigger model.

---

Implementation checklist for founders and developers

AX-style orchestration works best when you build the workflow around business constraints, not model novelty. Use this checklist before deploying.

1. Define the unit of work

Pick one repeatable task: one ticket, one company record, one pull request, one invoice, one research memo. Cost and quality metrics should be measured per unit.

2. Separate stages by skill

Do not combine classification, retrieval, reasoning, writing, and evaluation into one prompt. Splitting stages improves observability and makes it easy to route cheap models to simple steps.

3. Use structured outputs

Every intermediate stage should return JSON or another strict schema. Freeform prose is acceptable only for final user-facing outputs.

4. Add confidence thresholds

Set numeric thresholds for retry and escalation. For example:

• confidence >= 0.85: auto-approve low-risk output
• 0.65-0.84: run evaluator or premium retry
• < 0.65: human review

5. Track cost per stage

Log token usage by stage, model, workflow ID, and outcome. This lets you identify waste. In many agent workflows, retrieval summaries and evaluator loops consume more tokens than the final answer.

6. Cache stable context

Knowledge-base articles, policy docs, product descriptions, and company profiles should be cached. Re-sending the same context into every run increases cost without improving quality.

7. Create a fallback ladder

A sensible fallback ladder for many workflows:

1. Cheap model first pass
2. Same model retry with stricter prompt
3. Stronger model retry
4. Tool refresh
5. Human review

Skipping straight from cheap model to human review leaves automation value on the table. Skipping human review entirely creates operational risk.

---

Risks, limits, and when not to use AX-style agents

Agent orchestration adds power, but it also adds complexity. Not every workflow needs it.

Do not use an orchestrator for simple tasks

If the task is a single classification, short rewrite, or basic Q&A call, a direct model API call is faster and cheaper. AX is most useful when the workflow has multiple stages, external tools, or quality checks.

Watch for hidden token growth

Agents often accumulate context across loops. A support workflow that starts with a 500-word ticket can become a 40,000-token run after account history, KB articles, logs, and evaluation traces are included. Summarize aggressively and pass only the context each stage needs.

Avoid autonomous execution for high-risk actions

For refunds, account changes, payroll, legal edits, production deployments, and customer-impacting actions, the agent should draft and recommend. A human or deterministic rules engine should approve execution.

Treat traces as sensitive data

Execution traces may include customer data, internal strategy, source code, credentials, or financial information. Apply retention limits, redaction, access controls, and audit logging.

Measure quality against human baselines

Do not rely only on model confidence. Track actual outcomes: ticket reopen rate, enrichment accuracy, review acceptance rate, research correction rate, and time saved. A workflow that is cheap but creates cleanup work is not cheap.

---

Frequently asked questions

What is AX, Google’s open agentic orchestrator?

AX is an open agent orchestration approach for building multi-step AI workflows with planner, tool-use, evaluator, and fallback stages. Its main value is helping teams move from single prompts to traceable pipelines where each step can be inspected, retried, and routed to the right model.

How much does an AX-style agent workflow cost?

Most practical workflows cost $0.02-$0.80 per run depending on token volume, model choice, and how many evaluator or retry stages run. A support ticket can be $0.02-$0.08, while a deeper research memo can be $0.25-$0.60 with a balanced model stack. Use AI Cost Check to calculate your own run volume.

Which model should I use with an agent orchestrator?

Use cheap models such as GPT-5 nano, Gemini 2.5 Flash-Lite, or DeepSeek V4.1 Flash for classification, extraction, and formatting. Use stronger models such as GPT-5.2, Claude Sonnet 5, or Gemini 3 Pro for planning, judgment, synthesis, and final outputs.

When is a premium model overkill for agent workflows?

A premium model is overkill when the task has a narrow schema, low risk, and easy validation. Examples include document type detection, entity extraction, CRM field normalization, deduplication, and first-pass summaries. Save premium models for ambiguous decisions, high-stakes outputs, and final executive or customer-facing responses.

What should I build first with AX?

Start with a workflow that is frequent, messy, and measurable: support triage, CRM enrichment, pull request review, invoice processing, or weekly reporting. The best first project has clear inputs, clear acceptance criteria, and enough volume to justify orchestration.

---

CTA: calculate your agent workflow before you ship it

Before you put an AX-style agent into production, price the workflow per run and per month. Model routing can turn a promising prototype into a scalable system, while sending every stage to a premium model can quietly make the unit economics fail.

Use AI Cost Check to compare model prices, estimate token usage, and test premium-versus-fallback routing. For deeper model selection, review GPT-5 vs Gemini 3 Pro, GPT-5 vs DeepSeek V3.2, or browse individual model pages like Gemini 3 Pro, Claude Sonnet 5, and GPT-5.2.

The winning pattern for 2026 agent apps is clear: orchestrate the workflow, trace every step, evaluate before execution, and route expensive models only where their judgment changes the outcome.