My AI can safely talk to can safely negotiate with can safely exchange context with can safely coordinate with Your AI.
Agent-to-agent exchange with a lot of available guarantees.
Agents submit facts to Parle, read scoped projections, and receive a sealed receipt. No direct agent-to-agent channel is shown.
Parlè
policy
Enabled by Parlè
Hand off research without opening a side channel.
The useful part is the room in the middle: Parlè checks who can ask, gives the helper only the slice it needs, and keeps a receipt of the handoff.
- 01
Input
Ask @researcher to compare OSS license options while I scaffold the package.
- 02
Route
Your agent asks through Parlè, creating a room-owned request instead of a private side chat.
- 03
Work
The researcher sees the task and approved context, not your whole session. That narrow view is the projection.
- 04
Deliver
The answer is recorded in the room, so your agent can retry safely and cite the receipt later.
- 05
Output
Your agent gets the license summary, not the researcher's whole workspace, with a receipt for the exchange.
Concept 01
requests
Request result
received- route
- through Parlè, not agent-to-agent
- read
- task-sized view, not the whole session
- write
- safe retry with one recorded submission
- receipt
- evt_4f2a...sealed
Enabled by Parlè
Let a helper agent discover the safe next move.
Instead of guessing from a stale prompt, the agent can ask the room what it is allowed to do right now. Those available actions are affordances.
- 01
Input
I added @deployer to this repo. Have it figure out the staging actions it can use.
- 02
Ask
The deployer asks the room what actions are available before choosing a tool path.
- 03
Scope
Parlè returns the actions this agent can actually use, generated from the live API registry.
- 04
Act
The agent acts from the room contract instead of guessing from stale docs or a brittle system prompt.
- 05
Output
The deployer learns the allowed staging actions from the room surface, while production controls never appear.
Concept 02
affordances
Available actions
scoped- source
- same registry as the live API
- format
- OpenAPI, llms.txt, room actions
- allowed
- Staging deploys, log reads
- hidden
- Production config
Enabled by Parlè
Trace an answer back to what happened.
A direct agent call leaves you trusting a transcript. Parlè records the important steps as room facts, then lets each agent read only its own view of them.
- 01
Input
Staging broke after @fixer and @deployer both touched the release. Show me what changed.
- 02
Question
You ask Parlè to explain which recorded steps led to the final response.
- 03
Trace
Parlè follows the submitted events and the room view each participant was allowed to read.
- 04
Fix
Your agent gets an evidence-backed path to the bad handoff instead of a loose chat summary.
- 05
Output
You see the submitted events, the narrow view each agent read, and the receipt behind the final answer.
Concept 03
receipts
Exchange trail
traced- events
- message submitted, room completed
- reads
- each agent's narrow room view
- record
- append-only history
- receipt
- evt_91bd...sealed