Build together
A useful next step for every role
A dependable agent takes more than implementation. Use these guides to make product decisions, design conversations, evaluate outcomes, and prepare a release your team can operate.
We said Rasa never expands api_key: ${VAR}. Its client does
Rasa 3.21.0.dev3 rejects api_key_env. A test spy on the litellm call shows api_key: ${NAME} arrived as the real key on 3.20.0 too, and which line expands it.
Read the guide →How to audit a document an AI agent generates
We audited an agent that derives documents from records. Its tests pass; the engine call path, free-text reasons and a replaceable risk warning do not.
Read the guide →Build an evaluation set that exposes failures
Build an agent evaluation set with explicit task slices, traceable data, factual assertions, and failure cases that a headline success rate would hide.
Read the guide →Choose an agent workflow worth building
Decide which workflow deserves an AI agent using a worked scope brief, a measurable outcome, and a boundary your product and engineering team can test.
Read the guide →Test passes but the real code path fails: derive the fixture
A hand-typed session made a handoff test pass while the agent never collected account_id. Build fixtures from the key list the running code reads.
Read the guide →Design a handoff a customer can trust
Design and test an agent-to-human handoff with explicit consent, a useful context packet, and recovery paths when the receiving team is unavailable.
Read the guide →Stop customers repeating themselves after a handoff
To stop customers repeating themselves after a handoff, build the human agent’s first screen from typed fields and test the promise on the agent’s own path.
Read the guide →Rasa 3.20.0 sent endpoints.yml's api_key_env; OpenAI refused
On Rasa Pro 3.20.0 an endpoints.yml model group sends api_key_env to OpenAI as a request parameter. With the health check on, OpenAI refused it; Rasa exited.
Read the guide →Why TTS failover returns after a Rasa voice agent restart
A redeploy clears the voice router’s memory of a failing vendor; it never fixes the vendor. Read the failure class and pick wait, reset, re-route or escalate.
Read the guide →Build your first Rasa agent
Run a small Rasa agent locally, inspect the tool behind its answers, and make your first change with one complete downloadable starter.
Read the guide →When the caller ID match is not the caller
We argue a caller ID match identifies a number, not whoever holds the phone. Decide what a voice agent says before the caller confirms, and after a wrong match.
Read the guide →Make an agent release decision with evidence
Turn agent evaluation results into a release decision with an explicit denominator, a worked decision record, and owners for rollout and unresolved risks.
Read the guide →Haiku returned nothing and Rasa logged facts_count 0
Rasa Mantle memory discovery on Anthropic: Haiku gave Rasa 3.20.0's fact extractor no tool call in 20 replays, even after Please continue.
Read the guide →endpoints.yml said Ollama. Mantle still picked OpenAI
On Rasa Pro 3.20.0, Mantle takes the references embedder from agent.yml references.embeddings. Left empty, rasa train uses OpenAI and logs no provider.
Read the guide →Measure a handoff before you claim it worked
How to measure AI agent human handoff success: count the desk questions a context package retires, and keep that apart from handle time and CSAT.
Read the guide →186 mutations killed, and the guard still broke
A guard suite reported 186 of 186 mutations killed and still passed a one-word break. Test for the mutation nobody authored, and count the result honestly.
Read the guide →Review what an agent is allowed to do
Review an agent’s permissions with a worked action register, evidence from the tool boundary, and named owners for authorization, handoff, and unresolved risk.
Read the guide →Roll out an agent with a working stop button
Plan an agent rollout with dependency checks, outcome signals, a verified fallback, and a rollback drill that proves who can stop an unsafe release.
Read the guide →Is an MCP integration a drop-in replacement?
Is an MCP integration a drop-in replacement? A green proof shows the instructions held. Audit where each tool gets its customer id before you scope the swap.
Read the guide →What evidence to demand before approving an agent action
A clean transcript and a YAML constraint prove little about an agent action. The evidence a risk reviewer should ask for, row by row, before signing off.
Read the guide →After login, a stale Rasa skill line asked for the PIN again
Rasa Pro 3.21.0.dev3 drops skill-level requires. A precondition passed the traveller, and a leftover prompt line re-activated authenticate in 8 of 10 runs.
Read the guide →Ready to implement? Follow the Rasa tutorials for working examples and code.