Speed leads.
Reduce the wait for a useful answer. Keep the user informed and able to interrupt while work continues.
Botpress research
The fastest
trustworthy
agent harness.
A lightweight agent.
Deeper work in the background.
Tokebak handles immediate work using prepared information. Its design keeps the conversation open while larger models and harnesses take on deeper tasks.
See how the pieces fitTrust is built in.
Trace factual claims to their evidence. Check that the evidence supports them. Detect errors and correct what reaches the user.
The architectureOne conversation.
One conversation.
Several ways to do the work.
The foreground agent answers, acts and delegates. Results pass through evidence and intent checks before reaching the user.
Proposed architectureSelect a component
Voice and text share a foreground agent, supported by prepared data, tools and background workers. Enable JavaScript to explore the diagram.
Lightweight means little mandatory work. Adding a capability should not make every interaction pay for it.
Implementation notes
The runtime uses Rust, SQLite and GPT-Live. General background harness delegation and complete claim verification remain to be built. Components in this map can share a process.
Make the query path smallDo the slow
Do the slow
preparation early.
Fetch, parse, enrich and index data before the user needs it. Answer from prepared local data when it is sufficiently current and authoritative for the task.
Indexing may live outside the harness. It has freshness deadlines, rather than conversational deadlines.
Outside the conversation
Source data→Prepare→Publish index
Query-ready knowledge + structured data
When the user asks
Question→Local query→Evidence
Search the prepared policy directly, without fetching and indexing it while the user waits.
A dated snapshot cannot establish live stock.
AnticipationPrepare while the user
Prepare while the user
is still asking.
Like DNS prefetch before a click, prepare a likely next step before it is needed. The agent still has to validate the final request before using the work.
Prepare a lookup during input, then reuse or discard it when the user submits. Enable JavaScript to step through the examples.
The latency atlasHow we reduce latency
How we reduce latency
and perceived waiting.
The other north starCheck that the evidence
Check that the evidence
supports the answer.
A correct source can still become a wrong answer. Carry evidence through every transformation, check the delivered claim, and make corrections explicit.
“Unused items can be returned within 30 days.”Worked example · source A
You have 30 days to return an unused item.Cites source A
Supported
The policy and answer both say 30 days.
Traceability tells us where a claim came from. Verification asks whether the evidence supports it.
Voice adds another transformation. The model’s spoken paraphrase can change a fact. Evidence supplied to voice is not proof of what the user heard.
Checking before speech takes time. Buffering can hold an answer for verification. Checking afterward allows earlier speech but requires a correction if an error escapes.