AI · Agents · 720-hour cycle
kaedax builds production AI agents and AI features — agent products, copilots, AI-powered workflows — in a fixed 30-day (720-hour) cycle. The evaluation harness ships before the agent: graded test sets, regression gates, refusal layers, and MCP tool contracts. We run our own delivery on agents, so we build what we operate.
Eval suite
412/412
green
Task success
96.8%
+1.2 vs v12
Refusal accuracy
99.1%
0 false-do
Cost / task
$0.041
−18% mom
[ §01 ] the problems we're hired for
Every AI demo looks magical against cherry-picked inputs. Production means adversarial users, malformed data, and edge cases at volume. We build the eval harness before the agent — graded test sets and regression gates that make 'it works' a measured claim.
Prompt tweaks and model upgrades silently regress behavior. Without evals in CI, you find out from your users. Our builds gate every agent change behind the same green-suite discipline as code.
An agent that confidently does the wrong thing destroys trust faster than one that asks. Scoping what the agent won't do — and proving it with tests — is where enterprise buyers decide. We've written publicly on this; it closes deals.
Agents are only as good as their tools' contracts. We build MCP-based tool layers with typed contracts, sandboxed execution, and audit logs of every tool call — the plumbing that separates agent products from agent demos.
[ §02 ] what ships
The shapes of ai & agents work that fit a 720-hour cycle. Every build ships with runbooks, monitoring, and 30–60 days of post-launch on-call.
The agent, the tool layer, the eval harness, and the operator console — shipped as one product.
Copilots, summarization, extraction — dropped into your codebase with evals gating every release.
Graded test sets, regression gates, accuracy budgets — demonstrable to enterprise buyers and auditors.
Typed tool contracts, sandboxed execution, per-call audit logs, rate and cost budgets.
[ §03 ] proof
AURORA
TALLOW
They shipped the agent product *and* the eval harness that proves it works. We demo the evals to enterprise buyers now — it closes deals.
[ §04 ] questions
We build the evaluation harness before the agent: graded test sets from the spec, regression gates in CI, and accuracy budgets the client signs off on. Every prompt or model change runs the full suite before merge — 'it works' becomes a measured claim, not a vibe.
Market quotes for production agent builds range $50,000–$250,000+ depending on tool-integration depth. kaedax prices per fixed 720-hour cycle, shared on the scope call. One cycle ships an agent product with its eval harness and operator console.
Defaults: Claude as the planner model, MCP for tool contracts, LangGraph only where explicit graph orchestration earns its complexity, pgvector for memory, Inngest or Temporal for long-running jobs. We have opinions, ship with them, and deviate only for real engineering reasons.
Yes — a contained AI feature in your existing codebase is the most cycle-friendly shape we take. Our AURORA engagement shipped exactly that: an AI feature inside an existing product, eval-gated, in one cycle.
Model-swap safety is built in: the eval harness is model-agnostic, so upgrading models means re-running the suite and comparing scores — not re-engineering the product. You own the harness, the prompts, and the agents outright.
A 15-minute scope call with a kaedax founder and the engineer who'd lead your build. We either fit your ai & agents product into 720 hours, or we tell you why not — same call.
▷ other verticals