Actspan Agent example walkthroughs
Understand a business problem, follow its execution flow, run the example and check the result against evidence. Each guide identifies who owns the decision, which files you need, expected outcomes and what has actually been tested.
Start here
Start with the governed refund walkthrough. The complete public OpenAI path demonstrates an automatic payment, a local human-review checkpoint and a blocked quote, with receipts and a sandbox ledger. Expected completion: explain why each outcome occurred and locate its evidence.
Install the CLI and public examples first. The current public native CLI is qualified on Apple Silicon macOS. Live hosted provider requests use your account and incur charges. The local inference examples require Ollama and an existing model; local compute still has costs.
Know which files you can download
The free public examples archive contains exactly these five native workspaces:
| Folder | Public path and current qualification |
|---|---|
openai/refund_agent |
$75/$250 approved/$750 with matching ledger receipts. |
anthropic/refund_agent |
Bundled Claude $75 automatic receipt check; the broader isolated-plugin proof is separate. |
openai/customer_service_agent |
Fixture context, user-input pause and reviewed reply. |
openai/data_enrichment |
Company research followed by structured summary. |
openai/planning_helper |
Real clarification question and resumed plan. |
Keep the archive's sibling directories together: Claude refunds share the OpenAI sandbox service. The native CLI and examples are separate downloads. The provider matrix distinguishes exact public artifacts, installed plugins, source and historical checks. A new guide or flowchart does not qualify another provider path.
All other examples below require a separate design-partner handoff with the matching files, Python packages and setup notes. They cannot be obtained by cloning the private product repository as a visitor. Do not substitute similarly named files from the free archive: some older workflows have different targets and tools. Ask your existing Actspan contact for the matching bundle. The public contact address is info@actspan.com; mailbox delivery remains unverified. Use your existing working Actspan contact for a handoff; do not treat an untested address as a completed support channel.
The problem behind the examples
A convincing model response does not establish permission, reviewer authority, factual correctness or a committed write. The examples make those boundaries visible, rather than treating a sentence such as “refunded” as proof.
Your backend continues to enforce customer identity, ownership and business rules. Actspan governs the configured agent path and records evidence for mediated execution. Where configured, output checks compare claims with completed tool results. After a write, inspect the business record too: rejecting a final answer does not reverse a committed action.
Choose a business problem
| Problem | Complete walkthroughs | Evidence to look for |
|---|---|---|
| A refund needs limits, review and a trustworthy receipt | Public OpenAI refund, isolated Claude refund, LangChain refund | Quote tier, review or denial, completed payment receipt and persisted ledger. Local LangChain $250 denies payment; it does not demonstrate human approval. |
| A support response needs context and review | Public support review | Identity-input checkpoint, four fixture lookups, review decision and exact reply. This example does not pay a refund. |
| An assistant should answer from tools or queue a handoff | Native OpenAI tools, LangChain support, local Ollama support | Allowed knowledge/order lookup, unavailable foreign record and actual local handoff ID. A queue entry sends no notification. |
| A ticket claim must match an authenticated backend write | LangChain over MCP, native hosted MCP | Customer authorization, committed receipt and fresh ticket readback. The hosted native path needs separate live qualification. |
| Research must have a visible source and handoff | OpenAI research, Ollama research | Successful source lookup, recorded facts, summary input and evaluations. A valid schema does not prove facts are true. |
| Approved facts need summarizing, without new research | Native Claude facts | Two actual Messages outputs and a faithful handoff. This is supplied-fact transformation, not web search. |
| A plan should wait for missing information | Planning helper | Structured triage, real follow-up question, saved thread and continuation. A question checkpoint is not an approval. |
| A changed agent needs connected release evidence | Release and evidence | Matching revision, evaluation, authorized approval and deployment identity in a compatible partner portal. |
Assisted integrations and legacy patterns
| Example | Business lesson | Setup and scope |
|---|---|---|
| Customer-service routing template | Route returns, retention and information requests; review the return proposal. | Partner fixture; fixed-pattern screening is not universal injection protection. |
| Grounded support chat | Return explicit answer/follow-up/escalate fields and source references. | Partner fixture; an escalation recommendation contacts no one. |
| LiteLLM fixed-response demo | Distinguish Actspan admission from a gateway virtual-key denial. | Separate complete Docker/Compose fixture; no paid model replies. |
| Live gateway setup | Prepare model routing and matching scenario fixtures. | Separate partner handoff, Docker and hosted-provider key. |
| Legacy/reference: Gateway screening and replacement proposals | Inspect screening, retention and replacement proposals. | Matching legacy fixture required; incompatible with the public support-review files. |
| Gateway research | Inspect the same two-step research handoff through a gateway. | Separate prepared live setup; historical/current qualification distinguished in the guide. |
| Gateway planning | Resume a missing-detail checkpoint through the configured gateway. | Separate live setup; current public planning proof does not qualify this gateway route. |
Run and inspect from one workspace
Each guide gives its actual target, environment and commands. For local examples, keep the selected directory and saved state available through continuation:
export ACTSPAN_PORTAL_DISABLED=1
actspan init workspace.yaml
actspan validate workspace.yaml
actspan runs
actspan inspect "<RUN_ID>" --trace --json
Set the required provider credential privately before a live request. Initialization and validation do not establish provider access. Use the actual run ID, not a model-written identifier. Evaluations can make additional model calls and sandbox writes; read their setup and acceptance checks before running them.
Successful blocked business outcomes, pending checkpoints and execution failures are different states. On published 0.2.0's direct-prompt receipt path, check nested execution status and recorded error as well as shell status. The fixture in the refund guide demonstrates that historical reporting limitation explicitly.
Source files and compatibility
Public source-file labels identify what a handoff must contain; unpublished files remain plain text rather than links to a private repository. The guides are the public instructions. Use a matching CLI/runtime and example set, retain your lock/version with the evidence and keep source corrections separate from the artifacts that have actually shipped.
For the declaration behind an example, read the workspace reference, policy scope and execution evidence. Supported schema fields, passing tests and live verification are distinct claims.