01
Plan
Decomposes the goal into steps and states what it needs before it spends money or touches a system of record.
Task graphs · explicit preconditions · stop conditions
AgentsAgents that know
Agents that know
when to stop.
An agent is only as good as the machinery around the model: the tools it may touch, the memory it keeps, the gates it must clear and the evidence it leaves behind. We build all five.
03Hawthorn AI Labs
agent://run/8f21c
$goal: reconcile Q3 invoices against POs
$plan · 6 steps · 2 require approval
$tool: erp.search_purchase_orders()
$tool: ledger.match(line_items)
$3 exceptions found · evidence attached
$gate: write access → awaiting human
paused at policy gate
01 / Anatomy
Five capabilities, in order.
Most agent failures happen at step four. The design work is deciding exactly where autonomy stops.
02
Reason
Selects evidence, weighs conflicting sources and records why it chose one path over another.
Evidence selection · conflict resolution · rationale logs
03
Use tools
Calls your APIs, databases and internal services through typed contracts with timeouts, retries and idempotency keys.
Typed schemas · MCP · rate limits · idempotency
04
Take action
Executes the change behind policy gates, with a human approval step wherever the blast radius justifies one.
Policy gates · approvals · rollback hooks · audit trail
05
Self-improve
Turns traces into evaluations and corrected examples — reviewed by a human before anything is promoted.
Trace review · eval generation · gated promotion
02 / Architecture
The layer cake we deploy.
Each layer is independently testable and independently replaceable. Nothing depends on the model behaving perfectly.
01 Experience
Chat / APIWorkflow triggersHuman review console
02 Orchestration
PlannerRouterState machineRetry & resume
03 Capability
Tool registryMCP serversRetrievalMemory tiers
04 Control
Policy gatesPermissionsBudget & rate limitsApproval steps
05 Evidence
TracesEvaluationsCost & latencyAudit log
03 / Where agents earn their keep
Boring, high-volume, rule-heavy work.
If a task is high-frequency, mostly routine and expensive to staff, it is a candidate. If it is rare and highly novel, it usually is not.
Back-office operations
Triage, enrich, route and resolve tickets across systems that were never designed to talk to each other.
Email + ticketERP / CRMApproval gates
Document-heavy decisions
Underwrite, review or assess a package of documents with every claim traceable to its source page.
ContractsClaimsCitations
Engineering operations
Investigate alerts, reproduce failures, draft the fix and open a PR — with a human merging, always.
ObservabilityCI/CDCode review
Research & intelligence
Continuously monitor sources, detect meaningful change and produce a brief with confidence and citations.
Multi-sourceDedupScheduled
Data quality loops
Find anomalies across heterogeneous feeds, justify each flag and hand the ranked list to the people who act.
Anomaly detectionRankingEvidence
Internal knowledge
Answer from your own corpus with permission-aware retrieval, refusal on out-of-scope, and links to the source.
RAGACL-awareRefusal
04 / Control
Design the control planeAutonomy is a dial, not a switch.
We set the dial with you, per action class — and make it impossible for the model to move it on its own.
- Typed tool contracts — an agent can only call what the registry exposes
- Least-privilege credentials scoped per task, never a shared service key
- Budget, rate and blast-radius limits enforced outside the model
- Deterministic policy gates for anything irreversible
- Human approval at configurable risk thresholds
- Full trace capture: prompt, tools, outputs, latency and cost
100%of runs traced end to end
3 tiersof memory: working, episodic, semantic
Alwaysa defined stop condition and rollback path
Agentic systemsGive the busywork
Give the busywork
to an agent.
Start with one workflow, instrumented and supervised. Expand only when the evidence says it is safe to.
Typical first response within 24 hours · NDA-friendly