Continuous tool loop
One instruction triggers up to 25 rounds of multi-step execution with no human relay.
Core engine
Autonomous agent runtime
One instruction, a continuous tool loop until the task completes. The heart of the system, built entirely in-house.
What it does
Most AI tools wait for you to feed prompts step by step. TradeMax Agent takes one instruction, then decomposes, selects tools and executes on its own — up to 25 consecutive rounds — until it produces a result or hits a guardrail. Reasoning streams live, every tool call is visualized, and finished runs replay step by step in the run history.
The runtime core is hand-written with no framework wrapper: intent recognition uses zero-cost keyword routing, long conversations compress into a memory file instead of blowing the context, and when a write operation is hit, the run pauses for confirmation and then resumes from the exact round it stopped at.
The chat box is now the whole company's workbench: drop a PDF, Word, Excel or CSV into the conversation and the AI reads it — a customer RFQ or spec sheet analyzed on the spot, scans via OCR, parsed and discarded, never stored. The AI's reach scales by role: a rep's AI carries only the minimal tool set for their own leads and mail, enforced server-side, so no rephrasing escalates privilege. And AI email goes through a two-step gate — recipient locked to the customer record, sent exactly once after human confirmation, so even repeated clicks never send twice.
Core capabilities
Four budget guardrails
Rounds, total tokens, cost ceiling and wall-clock — any trigger stops the run with an explanation.
Resume from breakpoint
Writes pause for confirmation, then the run continues in place with context, cost and plan intact.
Feature checklist
- 76 real tools load conditionally on credentials — modules without keys never enter the registry, so hallucinated calls are impossible
- Every run is archived: tool-call arguments and results, per-call LLM cost, termination reason — fully replayable
- Run dashboard tracks success rate, average rounds, average cost and termination-reason distribution
- Conversations past 20 turns summarize in the background into memory, never blocking the chat
- Dual-provider streaming with automatic failover; every attempt is metered
- Chat reads files: PDF, Word, Excel, CSV, txt uploaded and read, scans via OCR; 15MB per file, up to 5 per turn, parsed then discarded — never stored
- The AI's reach scales by role: a rep's AI carries only a minimal tool set for their own leads and mail, enforced server-side and immune to prompt injection; sessions bind to the account, others' sessions refused
- Two-step AI send: the recipient is locked to the customer record — the model can't set it — and after confirmation the draft sends exactly once, so repeated confirms and concurrent clicks never resend
Want to see this module run?
Book a demo and we'll run this module on real data.