01 · TOOL USE
Claude returns well-formed tool calls — every time.
Production agents loop. Loop means tool_use → execute → tool_result → tool_use again. If 1 in 50 tool calls is malformed, you spend the next month writing parsers and retry logic instead of shipping features. Claude Sonnet 4.6 is the most reliable model we've tested at parallel tool calls + structured output across long sessions.
In production: Our recruiter agent makes ~12 tool calls per candidate (search → enrich → score → draft email → log CRM). In our internal eval set, Claude Sonnet returned a malformed tool call at well under 0.1%, handled with a single retry. Earlier-generation models in the same eval hit roughly 6× that rate.
02 · MCP
MCP is the agent-tool standard. Anthropic invented it.
Model Context Protocol is the open standard that lets any AI client call any tool server. Cursor, Zed, Continue, our backend — they all speak MCP. That means a customer with an internal MCP server for their CRM can plug it into our agents without custom integration work. The market is converging on MCP. Building on it is a no-regret move.
In production: Our backend exposes an MCPConnector that wraps any MCP server (JSON-RPC 2.0). Today: 12 first-party MCP integrations (Gmail, QuickBooks, Slack, etc). Tomorrow: customer-registered MCP servers become first-class agent capabilities.
03 · PROMPT CACHING
90% cost reduction on the parts that don't change.
Every agent run starts with the same ~3K-token system prompt, tool definitions, RAG context, and few-shot examples. Without caching, you pay full input rate on all of it, every run. With Claude's prompt caching, the cached portion costs 10% of normal — which is what gets a run cost down to $0.003. No other frontier model offers caching at this granularity.
In production: On the AP bookkeeper agent, prompt caching saves us roughly $0.01 per run at production volumes. That margin is what makes $0.50-per-invoice outcome pricing work without taking a loss on the long tail.
04 · SAFETY
Constitutional AI matters when buyers are CFOs and CIOs.
Regulated buyers — finance, healthcare, EU enterprises — read your model card. Claude's constitutional approach, low rate of confabulation on structured business tasks, and conservative default behavior on ambiguous instructions are reasons we close deals our competitors don't. It's not just a checkbox — it's why our agents refuse to send the email when the input is suspicious.
In production: On invoice processing, Claude flags ambiguous vendor-name mismatches as 'needs human review' significantly more often than earlier-generation comparison models in our internal eval. Fewer silent miscategorizations. Fewer post-deploy fires.
05 · EXTENDED THINKING
Sonnet 4.6 reasons before it acts. That changes the failure mode.
Extended thinking lets the model produce internal reasoning before its final answer. For agents, that means fewer 'confidently wrong' tool calls. We use it on the bookkeeping agent before posting to QuickBooks, on the recruiter agent before sending an outreach email, and on the voice agent before classifying a call outcome.
In production: Adding extended thinking on AP categorization measurably cut our human-override rate in internal evaluation — meaningfully shrinking the approval queue per customer per day. We publish a current methodology + numbers under NDA.
06 · OPUS WHEN YOU NEED IT
Three model sizes, one API, one mental model.
Haiku 4.5 for routing ($1/M in, $5/M out). Sonnet 4.6 for reasoning ($3/M, $15/M). Opus 4.7 for the hardest tasks ($15/M, $75/M). Same API, same tool format, same safety properties. We route between them inside a single agent run — cheap classification followed by expensive reasoning followed by cheap formatting — without any infrastructure switch.
In production: Our recruiter agent uses Haiku to filter 500 candidates down to 20, Sonnet to rank and personalize outreach for those 20, and Opus only on the 2-3 'need a really thoughtful intro email' candidates. Effective per-candidate cost lands in the low single-digit cents at production volume.