// 01
Planner agents
Top-of-graph agents that decompose intent into a working plan, assign work to specialists, and re-plan when the world changes mid-execution.
- Goal decomposition
- Re-plan on tool failure
- Cost-aware tool selection
Top-of-graph agents that decompose intent into a working plan, assign work to specialists, and re-plan when the world changes mid-execution.
A second pair of eyes that disagrees with the planner. Tonal mismatches, missed constraints, brand drift — caught before the artifact lands.
Specialists with bounded scope, deterministic side-effects, and clear contracts. Tools are the only way agents touch the world.
Scoped, observable memory. Cross-session continuity where it matters; forgetting where it does not. No black-box state.
We map the operation as it runs today — actors, decisions, the moments humans actually own. Agents go where the cost of judgement is high and the work is repetitive.
Who plans, who critiques, who acts on the world. We name every agent, write its contract, and decide what it cannot do — before it ships.
Python · LangGraph-class orchestration · structured tool I/O · evals from day one. Observability is not an afterthought; it is the first thing we build.
A named human owns the loop. They can read every decision, override any output, and reverse anything we shipped. That surface is the product.
That works for demos. It does not work for operations.
A chat box is one agent, no critic, no memory boundary, no audit trail, no reversal path. The first time the model is wrong in production — which it will be — there is no surface to read what happened, fix it, and prevent it from happening again. We have inherited engagements that started as chat boxes; the rebuild costs more than starting clean.
We design around them, not against them. The critic agent's job is to catch them; the tool contracts make them visible; the operator can roll back any artifact a hallucination produced.
The model will say untrue things — we treat that as a property, not a defect. The system either catches it before it acts, or makes it reversible after.
The operator is the named human who lives inside the loop. They are not the buyer or the executive — they are the person who answers the phone when the system surprises someone.
The operator surface is the product. Every decision the agent makes is legible to them. Every output is reversible by them. Every constraint they tighten propagates back into the critic. We do not start engagements without one named.
Yes. We deploy to AWS, GCP, or on-prem. We have shipped systems into private VPCs, into government-cleared environments, and onto bare metal in a Norwegian basement. Where your data lives is your choice.
Minimum six weeks. Typical 8–14 weeks to first stable production deployment, then a quiet retainer for evolution.
We do not run two-week pilots. They produce demos, not systems.
We have inherited several. The first two weeks of an inherited engagement are spent reading the existing system; the read is itself a deliverable. Sometimes we extend, sometimes we replace. We say which, in writing, before we charge to build.
Send a brief. A principal reads it within 48 hours and replies with an architecture sketch — yes, no, or referral. We onboard two systems per quarter.