Workflow template gallery¶

Pick one team task, run its fixture, then adapt its inputs and tools. Each walkthrough takes about five minutes after the quickstart. All five run in the same Compose trial without credentials for any external service, a GPU, or framework extras.

Team task	`--template`	What you can inspect
PR review with human approval	`code-review`	Findings, a waiting draft, and an authenticated approve/reject decision
Support ticket triage	`support-triage`	Category, priority, suggested owner, and a reply draft
Weekly report from several sources	`weekly-report`	Changes, support and incident snapshots alongside the report
Incident summary from logs	`incident-summary`	UTC timeline, log references, mitigation and open questions
Document Q&A with citations	`document-qa`	Answer, source links and the exact retrieved excerpts

Before you start¶

Complete quickstart step 1 from this checkout, keep the fake model route, and return to the checkout root. Keep AGENTWORKFLOWS_API_KEY=local-development-only set in your shell. Every command below works in Bash and PowerShell. Replace RUN_ID with the UUID returned by runs start; repeat inspect while a run is still working. The console at http://127.0.0.1:8080/console/ offers the same workflows and sample inputs under Run workflow.

init works offline and refuses to overwrite a nonempty directory. It writes workflow.py, worker.py, input.json, a pinned SDK requirement, .gitignore, and local instructions. The built-in worker already runs these unchanged examples. Input edits apply to the next run immediately; source edits require your own worker.

The fake returns canned answers and synthetic data, even if you change the input. It proves the workflow, policy, approval and receipt paths; it does not evaluate model quality or connect to your GitHub, ticketing, logging, or document system. The example.test source links identify fixtures and are not live pages. All examples request the existing default budget of 10,000 tokens / $5, further limited by team policy; fixture prices are synthetic.

PR review with human approval¶

For a developer who wants a reviewable security/correctness draft before acting on a PR:

agentworkflows init my-review --template code-review
agentworkflows runs start CodeReviewWorkflow --input '@my-review/input.json'
agentworkflows runs inspect RUN_ID

The input's diff contains a patch to auth.py. At progress.stage: awaiting_approval, read progress.draft: the fixture flags the unconditional True at auth.py:10, suggests restoring the admin check, and asks for admin/non-admin tests. Then submit your decision:

agentworkflows runs approve RUN_ID
agentworkflows runs inspect RUN_ID

Expect status: completed, result.approved: true, the review and reviewer, and model and approval receipts. Start another run and use runs approve RUN_ID --reject to see result.approved: false. Both decisions finish the run; neither posts, executes or merges code. Approval expires after seven days. The trial's Slack/webhook/email alerts go to the local notification inbox.

Adapt it: put a sanitized unified diff in my-review/input.json. Include file names and hunk line numbers so findings can be checked. Use separate builder and approver identities from team setup. A PR webhook needs an adapter that fetches the diff and submits {"diff":"..."}; this template does not accept a raw GitHub event. If you add a comment-posting tool, call it only after a true approval and make the endpoint honor the gateway's idempotency key. Check findings against the actual diff.

Support ticket triage¶

For a support lead who needs routing and a reply draft while keeping customer communication under the team's control:

agentworkflows init my-support --template support-triage
agentworkflows runs start SupportTriageWorkflow --input '@my-support/input.json'
agentworkflows runs inspect RUN_ID

The input's ticket describes three teammates unable to sign in after password resets. Expect status: completed and a text result with category access, priority high, owner identity-support, and a draft asking for the error and time without requesting secrets. There is one model receipt and no customer message is sent. No approval is required to create this internal draft.

Adapt it: replace ticket with a sanitized ticket body. Edit the prompt to use your actual queues, severity definitions and escalation criteria. Connect a ticket-system adapter through the signed webhook contract, normalizing input to {"ticket":"..."}. Check suggested priorities against known tickets before routing live work; add human review before any customer-facing send action.

Weekly report from several sources¶

For a team lead assembling a briefing from engineering, support and reliability:

agentworkflows init my-weekly --template weekly-report
agentworkflows runs start WeeklyReportWorkflow --input '@my-weekly/input.json'
agentworkflows runs inspect RUN_ID

The input's period is 2026-09-28/2026-10-04. The workflow calls report_source separately for changes, support, and incidents, then asks for a single report. Expect status: completed, result.sources with three snapshots, and result.report linking each source. The fixture reports two shipped changes, 12 opened / 9 closed tickets and INC-1042; it explicitly flags the missing backlog baseline. The timeline has three tool receipts and one model receipt. The report stays in the run result.

Adapt it: implement the existing report_source POST contract, accepting source and period and returning {id, url, period, text}. Point its team tool URL at an authenticated adapter for your repositories, helpdesk and incident tracker. Bound date ranges and response sizes, and use stable source links. A failed source call fails the run after retries; this template does not silently turn missing data into a successful report.

For a Monday 09:00 UTC run, merge this into the existing WeeklyReportWorkflow policy in deploy/compose/sandbox-policy.yaml, preserving its models, tools, egress and budgets:

triggers:
  weekly:
    kind: cron
    project: default
    cron: "0 9 * * 1"
    input: {period: previous week}

Recreate the gateway from the checkout root (docker compose -f deploy/compose/compose.yaml up -d --force-recreate inference-gateway; prefix with wsl.exe -d Ubuntu -e on Windows). Your real adapter must resolve previous week into a UTC range; the fake always returns the same synthetic snapshots. Inspect and pause the schedule after trying it:

agentworkflows triggers list
agentworkflows triggers pause WeeklyReportWorkflow weekly

Use triggers resume WeeklyReportWorkflow weekly when ready. The manual run above exercises the same report immediately, so there is no need to wait until Monday. See schedule semantics for overlap and catch-up limits.

Incident summary from logs¶

For an incident lead turning a bounded log export into a draft handoff:

agentworkflows init my-incident --template incident-summary
agentworkflows runs start IncidentSummaryWorkflow --input '@my-incident/input.json'
agentworkflows runs inspect RUN_ID

The input's incident_id is INC-1042. The incident_logs tool supplies four timestamped log lines. Expect status: completed, result.logs retaining those lines, and result.summary citing [L1] through [L4]: deployment at 09:00 UTC, errors at 09:05, rollback at 09:12, and a return to baseline at 09:20. The fixture labels the deployment as an unconfirmed cause and asks for the affected-user count. There is one tool and one model receipt. It performs no remediation and declares no incident closed.

Adapt it: replace the incident_logs endpoint with a read-only log adapter accepting {incident_id} and returning {incident_id, source, lines}. Enforce the team's access, redaction, time window and size limits there. Keep UTC timestamps and stable log IDs. The incident commander should verify the draft against logs and metrics before sharing it. Raw logs and results enter Temporal history; receipt redaction alone does not remove that data.

Document Q&A with citations¶

For an internal helpdesk answering questions from a small, approved document collection:

agentworkflows init my-docs --template document-qa
agentworkflows runs start DocumentQAWorkflow --input '@my-docs/input.json'
agentworkflows runs inspect RUN_ID

The input's question asks who can approve a workflow and when approval expires. The search_documents tool retrieves a handbook excerpt. Expect status: completed, result.answer mentioning admins/approvers and seven days with [S1], and result.citations containing S1's title, URL, and exact excerpt. The timeline has one retrieval tool receipt and one model receipt.

Change the question in my-docs/input.json to What is the weather on Mars? and start a new run. The fixture retrieves nothing: expect I don't know from the supplied documents., empty citations, and no model call. Fixture retrieval recognizes approval and receipt questions only; it is independent of the optional RAG service in Compose.

Adapt it: connect search_documents to your approved search/RAG backend. It accepts {query} and returns a list of {id, title, url, text} excerpts, each with a unique stable ID and source link. Enforce document access in that adapter and limit excerpt sizes. The workflow requires JSON model output and checks that inline citation IDs match the returned IDs and retrieved excerpts. Malformed JSON or fabricated/mismatched IDs fail the run with a non-retryable error; an empty citation list yields the same abstention. Review the error and source before starting a corrected run. Valid IDs do not prove the cited text supports every claim; evaluate answer quality and abstention on your corpus.

Research starter¶

The original default remains available: agentworkflows init my-research. Follow the quickstart to run ResearchWorkflow with a topic, inspect its two model steps, approve its draft, and see the local publish fixture. It also accepts token_limit and cost_limit_usd.

Notifications and next steps¶

All template inputs accept an optional model, defaulting to demo-openai. Supply the nonempty text field shown in each walkthrough; unknown fields are rejected before Temporal starts the run. agentworkflows usage shows estimated spend across runs. Inspect each timeline for receipt IDs; use the quickstart's receipt export to verify the retained chain.

Compose sends approval-waiting, failure, and budget-threshold notifications to a local fake inbox at http://127.0.0.1:8025/ (JSON). Nothing goes to a real Slack workspace or mailbox. Completed reports are run results, not automatic outbound report deliveries. Triggers and notifications explains setup for your team's authorized destinations.

Before real use, configure approved tools/models, separate team credentials and budgets, and evaluate model output. Sanitize input before submitting it: workflow inputs, activity results and drafts are retained in Temporal even though gateway receipts fingerprint prompts. See production readiness. When finished with the trial, follow Stop the trial to remove its containers and volumes.

Run your edits¶

The Compose worker runs the installed examples. To execute your edited workflow.py, stop that worker from the checkout root before starting yours:

docker compose -f deploy/compose/compose.yaml stop workflow-worker

On Windows, prefix Compose commands with wsl.exe -d Ubuntu -e. In a second terminal, activate the same SDK environment as in the quickstart, enter the generated project, and start the worker:

BashPowerShell

export AGENTWORKFLOWS_API_KEY=demo-worker
python worker.py

$env:AGENTWORKFLOWS_API_KEY = 'demo-worker'
python worker.py

Wait for Worker ready on demo-workflows. Start and inspect runs from the original terminal using its human credential. Your worker registers only this project's workflow; do not start other templates on that queue until restoring the built-in worker. Restart your worker after changing source. Stop it with Ctrl+C, then restore Compose:

docker compose -f deploy/compose/compose.yaml up -d workflow-worker

If you chose a different gateway port, set AGENTWORKFLOWS_URL in the worker terminal too. If you chose a different Temporal port, set TEMPORAL_ADDRESS=localhost:PORT there. For your own team, set AGENTWORKFLOWS_TEAM and its worker key; the queue defaults to TEAM-workflows. Register the workflow name, approved models/tools/egress, and budgets in the existing team policy.

A workflow in about 20 lines¶

This is the approval pattern used by the code-review template, with typed input:

from dataclasses import dataclass
from temporalio import workflow
from agentworkflows.workflows import ApprovalWorkflow, WorkflowGateway

@dataclass
class CodeReviewRequest:
    diff: str
    model: str = "demo-openai"

@workflow.defn
class CodeReviewWorkflow(ApprovalWorkflow):
    @workflow.run
    async def run(self, request: CodeReviewRequest) -> dict:
        review = await WorkflowGateway().text(
            f"Review this diff for correctness and security. Cite lines and suggest tests.\n{request.diff}",
            model=request.model,
        )
        approved = await self.approval(review)
        return {"approved": approved, "review": review, "reviewer": self.reviewer}

The approval base implements the signal, query, update, and expiry behavior used by the run API. text schedules the same governed model activity as the lower-level model method. See SDK reference for defaults and bring your agent for OpenAI, Anthropic, Agents SDK, LangGraph, MCP, and sandboxed code steps.