You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$21,000 left to reach the funding goal.
When several AI agents work on the same tool at once, such as a code repository, a shared document or a reference library, they overwrite each other and often nobody notices. We measured this on live services using scratch data only. With 4 to 16 uncoordinated writers on one item, the Zotero API rejected 51 to 92% of writes as stale. Google Drive has no conditional write, and 70 to 95% of updates that it had confirmed were silently lost.
We have designed a fix that comes from mobile radio networks. A small coordinator gives each agent a short, exclusive time slot at each tool, with a fencing token. When the slot is granted, the coordinator checks that what the agent read is still current, so a plan built on stale data is rejected before it writes. The design is in a paper under journal review, with a public preprint available on request (project page: https://sonata42.com/slot-gateway/). In 17,640 simulated runs with 2 to 64 agents there were no deadlocks and no stale writes. At 16 agents the 95th-percentile task time was 160 seconds, against 1,731 seconds for a global lock.
All of that is simulation. This project takes the first step out of the simulator. We will release the code, build the gateway as a Model Context Protocol (MCP) server, and test it with real language-model agents on real tools. Everything will be open source under Apache 2.0.
Goal 1. Release the simulator and measurement code so others can check and reuse it.
Goal 2. Build Slot Gateway v0.1, an MCP server that sits in front of a shared tool and hands out slots and fencing tokens.
Goal 3. Run the first public tests with real language-model agents, comparing four set-ups on the same tasks: no coordination, a global lock, optimistic retries, and the slot protocol.
With the minimum (2 months):
Public code release (simulator, measurement scripts, gateway v0.1) under Apache 2.0.
Adapters for a local Git repository and a test file store.
A public test plan, written before data collection.
A pilot with 2 to 8 real agents on a scratch Git repository, with a short public results note.
With the full amount (4 months), in addition:
Adapters for Zotero and Google Drive, the two services where we measured the losses.
Real-agent tests with 2 to 16 agents on all four tools, with the full comparison above.
A first version of provenance labels in the gateway, so a value an agent reports can be traced to the tool read that produced it.
A public dataset and a short preprint with the results, including any results that go against us.
How: the co-lead wrote the protocol and the simulator and will build the gateway and run the measurements. The lead brings verification and agent-coordination tooling and research on reliable multi-agent behaviour. We use AI coding agents daily to write adapters and tests, and the experiments use language-model agents as subjects. All tests use scratch accounts and scratch data only.
With the gateway in place, no stale or lost writes on any of the four tools, against measured loss rates for uncoordinated agents on the same tasks.
95th-percentile task time lower than a global lock at 8 and 16 agents.
Reasoning tokens thrown away, reported honestly against optimistic retries. The simulator says the slot protocol loses to optimistic retries at 32 to 64 agents, and we will report whether that holds with real agents.
Someone outside the team can install the gateway in front of a Git repository and reproduce our main result from the published code and instructions.
Researcher time and API costs only. The grant is paid to Sonata42 Limited.
Lead, about 2 days a week, USD 3,000 per month: 6,000 (minimum, 2 months) / 12,000 (full, 4 months)
Co-lead, about 1 day a week, USD 1,500 per month: 3,000 / 6,000
Language-model API use for real-agent tests: 1,000 / 2,000
Overhead 5% (company accounts, bank, archive hosting): 500 / 1,000
Total (USD): 10,500 / 21,000
This is the first part of a larger 12-month proposal now pending with the Foresight Institute. We are also applying to other funders for the same pilot. If more than one funds, we will tell each funder and run the funded pieces one after another, so no month of anyone's time is paid twice.
The project lead is a former university senior research fellow in informatics. The lead builds agentic AI systems and a typed language for agent coordination with a formally verified compiler, has published on institutional design for reliable multi-agent behaviour, has led research grants as principal investigator, and has built and run internet routing (BGP) networks since 1997.
The co-lead has about 20 years in telecom and network engineering and wrote the slot protocol and simulator.
We work together through Sonata42 Limited, a two-director AI research company in Edinburgh (company SC797083).
Real agents behave differently from the simulated ones, for example much longer or more variable reasoning times, so slots are held too long and the gain over a global lock shrinks. Outcome: we publish the measurements and the reasons, which still tells others where the simple fixes break.
Some tools have no conditional write (Google Drive is one), so the gateway has to keep its own version check. If that cannot be made reliable, we will say so and limit the claim to tools that support it.
The token waste at high agent counts holds up with real agents. Outcome: we report it and a hybrid mode becomes the next piece of work.
Team time. With only two part-time people, illness or other work could slow the project. If that happens we will extend the timeline rather than cut the testing, and return any funds not used.
In each case the code, data and a written report are still released.
None. A 12-month proposal for the wider version of this work is pending with the Foresight Institute AI for Science and Safety Nodes programme (submitted November 2026, decision expected January 2027). We are also applying to other AI safety funders. If more than one funds, no month of anyone's time is paid twice.
There are no bids on this project.