Step-by-step, code-heavy guides for building AI agents.
84 posts
The beeai-framework docs say onlySuccessInvocations affects forceAtStep. It also silently scopes maxInvocations, consecutiveAllowed, minInvocations and onlyAfter. Measured end to end on 0.1.31, with a control arm on every group.
If your agent has a tool that takes a URL, the model chooses where your server connects. Measured end to end on Node 24 with no dependencies and no API key: the guard most people write fails four ways, including the IPv4-mapped IPv6 bypass behind CVE-2026-56266, which also defeated the first version of my own hardened guard.
On @langchain/langgraph 1.4.19 the package root's type declarations export 212 names and its runtime exports 95. Ten of the gap are classes, tsc compiles the broken import without complaint, and one of the imports that fails is copied verbatim from the vendor's own JavaScript documentation.
ConditionalRequirement accepts onlyBefore, validates it five ways, and never consults it. A seven-arm harness measures the tool set a BeeAI RequirementAgent actually hands the model.
On @inngest/agent-kit 0.13.2 the option maxIter exists on both Agent.run() and createNetwork(), both default to 0, and 0 means exactly one inference in the first and no limit at all in the second. A harness that measures both with no API key.
The prompt injection guardrail is third in the bundle and the HTML sanitizer is fourth, so markup the sanitizer strips rejoins a banned phrase after the detector has already passed it.
On @langchain/langgraph 1.4.19 a DeltaValue field replays checkpointed writes through your reducer in larger batches than your graph produced. If the reducer reads the batching, the resumed state is not the state you ran, and nothing errors. Three runnable scripts: a ten-line invariance check, a run where six turns become one, and a sweep showing the damage sits in the tail.
A misspelled path in a Mastra workflow branch predicate never errors. On @mastra/core 1.74.0 five of nine operators then skip the branch and four run it, so ne and not wrapping eq disagree on the same question. Four runnable scripts print the table, and the checker that would have caught it ships in the same package wired elsewhere.
A suspended workflow can only resume if its state was written somewhere. On @voltagent/core 2.11.0 there are three states, not two: no workflowMemory fails even in the same process, the in-memory adapter dies at process exit, and only a durable adapter survives a restart.
On @openai/agents 0.18.0, the tool calls of one turn share a cancellation signal. One fatal failure aborts every other tool in the turn with SiblingCancellationError, and a tool that ignores details.signal keeps running, so its side effects land after the run has already thrown.
A Mastra output processor that implements only processOutputResult does run on agent.stream(), but it runs after every chunk has been delivered. The redaction lands on result.text and never reaches the reader. Measured on @mastra/core 1.72.0 with no API key.
Measured on @openai/agents 0.18.0: the built-in client-side tool_search loader matches paths against tool names by exact string equality, and a miss returns zero tools with no error and no warning. The failure arrives one turn later as an error that blames the model.
