An engineering organisation you can ask questions

Call sites never name a model. They request a use (lead, fast or review) and a per-thread catalog resolves it. A thread pins one vendor for its whole life, so protocol-level behaviour stays consistent even as the transport underneath changes. Adding a new role costs a catalog entry, not a rewrite of provider construction.

Complex research runs as a three-stage workflow. Sonnet 4.6 plans the investigation and identifies the evidence required. Haiku 4.5 performs high-volume tool selection and data retrieval. Sonnet then reviews the final result, checks whether the question was actually answered, and sends incomplete or unsupported work back for revision.

Tool calls run in parallel when their results are independent. Long investigations maintain a task list, compact context as the conversation grows, and validate citations before returning an answer, which is what moves Src past retrieval into sustained investigation without losing the link between conclusion and evidence.

Claude is reachable two ways, and Span uses both. A Redis-backed policy watches for rate limits and upstream errors on the primary; past a threshold inside a window it flips the fleet to the secondary transport for a fixed duration, then flips back.

The distinction that matters: the transport changes, the vendor does not. The same Claude model answers, so nothing about the thread's behaviour shifts under the user mid-investigation.

The agent calls Claude's code_execution server tool alongside its own local tools, so it can analyse large result sets without putting every row into the conversation context. That combination has semantics worth knowing: a turn calling both returns the local tool calls plus a pending server_tool_use block, and the code execution result only lands on the next call, after the local ToolMessage is submitted. Span mapped that two-step pattern explicitly rather than discovering it in production.

Long investigations exceed any window. When a thread approaches its limit, older history is summarised by a separate model call while unanswered user messages and recent turns stay literal, then reassembled behind the original system prompt.

The empty-message and whitespace rules the Messages API enforces are documented from live testing and filtered before the request goes out, one fewer class of 400 in production.

A dedicated Claude-powered coach analyses each coding session at three levels: the overall trace, individual user turns, and the underlying events, meaning prompts, tool calls, file reads, edits and sub-agent activity. Claude calibrates its analysis against the type and complexity of the task, distinguishing healthy exploration from unnecessary tool usage, and effective autonomy from incomplete validation.

The rule that keeps it useful: a recommendation must be supported by specific, quoted behaviour rather than an aggregate score. That is the difference between coaching and a productivity ranking.

Usage metadata from each call feeds a ledger that separates input, output, cache reads and cache writes, prices them per model, and applies the regional multiplier when the call went through Bedrock. That ledger backs a per-organisation monthly budget in dollars with ok / warn / blocked states.

Because the budget is enforced at the call rather than reconciled at month end, cost becomes a product guarantee instead of a surprise.

The metrics specialist translates natural-language questions into queries against Span's engineering knowledge graph, connecting numbers to the teams, people, services, pull requests, issues, deployments, incidents and workstreams around them. Further downstream, Claude classifies engineering work by type across an entire organisation's history, and those labels roll up into the capex/opex split the finance side reports on.

This is the part worth being precise about: model output here is not advisory, it is the number on the report. A classification that feeds a financial report has to survive review, so sampling, an audit trail back to the source commits, and a documented prompt version are not extras. They are the control.

Src uses Claude to replace static engineering reporting with an interactive understanding layer. Teams explore code, investigate operational questions, build leadership reports, analyse AI effectiveness and understand engineering behaviour through one source-backed conversation.

Every conclusion is designed to show its evidence, which makes the result something teams can verify and act on, rather than another opaque AI summary.