Stories from people dealing with Tool Use and Function Calling — what changed, who decided it, and what the ordinary days look like now.
20 stories
Agentic AI & Orchestration
A timed-out tool call can leave an agent retrying an order that already exists. Recovery depends on treating the result as unknown until the remote system confirms what happened.
Mara Quintero · October 3, 2026 · 8 min
AI Industry & Models
OpenAI can now retain a model run’s messages and tool context. That reduces orchestration code, but it makes deletion, storage duration and provider portability part of the API design.
Tobias Lund · 8 min read
Agentic AI & Orchestration
A vendor support page is untrusted input, even when an agent needs it to finish a task. This walkthrough tests whether page text can trigger data leaks, unsafe tool calls, or account changes.
Mara Quintero · 8 min read
Agentic AI & Orchestration
An ordering agent can retry after a timeout and buy the same item twice. A durable idempotency key lets the order and payment systems recognize one purchase intent across repeated calls.
Mara Quintero · 8 min read
Agentic AI & Orchestration
A successful tool call can leave the requested job unfinished. The reliable check reads the systems that own the order, warehouse task, or other business outcome.
Mara Quintero · 8 min read
Agentic AI & Orchestration
A successful API response can carry empty, stale, or incomplete data. Test whether an agent notices the defect before grading whether it finished the task.
Mara Quintero · 8 min read
Agentic AI & Orchestration
A browser agent can carry a hostile download from a public site into email or cloud storage. Put an inspection gate between the download tool and every upload tool.
Mara Quintero · 7 min read
Agentic AI & Orchestration
A timed-out tool call can leave an agent between failure and success. Safe retries depend on a persistent request identity, a way to check status, and a queue for unresolved actions.
Mara Quintero · 7 min read
Agentic AI & Orchestration
A natural-language database agent can handle recurring analysis without arbitrary SQL access. The workable setup combines narrow views, enforced query budgets and a separate path for changes.
Mara Quintero · 8 min read
Agentic AI & Orchestration
A customer-service agent should not carry one reusable credential for every support action. Split access by action stage, then issue narrow, temporary credentials only when execution is approved.
Mara Quintero · 8 min read
AI Industry & Models
A code sandbox can turn a plausible answer into a reproducible calculation. It also needs strict limits on network access, files, packages, runtime, and output.
Tobias Lund · 8 min read
AI Industry & Models
Basic chat requests transferred unchanged in our test. Parallel tools, strict argument schemas, streamed calls and tool results exposed differences that require an adapter.
Tobias Lund · 7 min read
Agentic AI & Orchestration
A parser error persuaded a sandboxed agent to read a canary token. Typed results and an external action broker stopped the same tool output from becoming an instruction.
Mara Quintero · 8 min read
Agentic AI & Orchestration
A browser agent followed instructions embedded in a vendor support page instead of retrieving an invoice. Isolation reduced the confusion; narrow tool permissions stopped the attempted data leak.
Mara Quintero · 8 min read
Agentic AI & Orchestration
Indirect prompt injection turns ordinary page content into instructions for an agent. The practical defense is to limit what the agent can see, call, and approve on its own.
Mara Quintero · 8 min read
Agentic AI & Orchestration
A browser agent followed an off-screen instruction into a dummy internal file. A URL allowlist and action-level confirmation gate stopped the same run without relying on the model to police itself.
Mara Quintero · 8 min read
AI Industry & Models
A retirement date tells you when an endpoint closes, not whether its replacement behaves the same. Shadow traffic exposes changes in refusals, tool calls, latency, and length before cutover.
Tobias Lund · 8 min read
Agentic AI & Orchestration
Indirect prompt injection turns page content into commands for an AI agent. Authenticated sessions raise the stakes because the agent may already have access to files, email, purchases, or account settings.
Mara Quintero · 8 min read
AI Industry & Models
Valid JSON is only the first gate. Production systems must also catch truncated responses, schema drift, unsupported fields, and values that look valid but trigger the wrong action.
Tobias Lund · 8 min read
Agentic AI & Orchestration
Replay historical support threads in an isolated sandbox, capture each tool call, and grade four kinds of behavior before the agent receives permission to send.
Mara Quintero · 8 min read
Other impacts
One story a day
One real story about AI each morning — no hype, no alarm, just company for the road.