Agent infrastructure · Measured Intuition
Giving AI agents shared context and verifiable handoffs
We built persistent shared records and version checks for work spanning multiple agents and conversations. The system preserves decisions, records handoffs and checks completion evidence against the current implementation.
The operating problem
When work spans multiple agents and conversations, decisions fragment. Specialists act on different assumptions, context gets repeated and a successful test can become outdated after the next change.
The challenge was to preserve continuity while making completion claims inspectable. Adding more agents increases the need for shared context, clear handoffs and current evidence.
What was built
Connected specialist agents through shared project context, reusable skills and explicit handoffs. Built a hosted MCP plugin with OAuth and SQL-backed persistence so work could be saved, retrieved and continued across conversations.
A Python/SQLite verification layer links decisions to requirements, code, tests and client-facing explanations. Checks bind to specific file versions. Relevant changes invalidate earlier evidence until it has been reviewed against the new state.
Separate interpretation from verification
- Agents interpret requests, investigate problems and propose changes.
- Persistent records preserve decisions, work state and supporting evidence.
- Code checks versions, duplicate writes and whether verification remains current.
- Independent review assesses whether implementation satisfies the intended behavior.
This structure allows work to be distributed without treating an agent’s success message as sufficient proof of completion.
Measured retrieval result
A bounded evaluation returned 96.9% less data for a scoped retrieval than a full-state read: 3,513 bytes compared with 113,598. The relevant work and evidence remained included.
Resuming unchanged work preserved record IDs and versions. Hosted persistence and readback were demonstrated.
This is a retrieval-volume result from an evaluation fixture. It is not a claim of equivalent token savings, lower operating costs or faster production delivery.
Business significance and current boundary
The system addresses coordination overhead: repeated explanations, conflicting context and rechecking work after changes. It supplies infrastructure for longer-running delegation with traceability and review.
Unattended hosted acceptance and some platform behaviors remain incomplete. The verification layer checks declared scope and current evidence. It does not establish that every possible requirement has been discovered or that human supervision is unnecessary.
Discuss a workflow worth changing.
Start with the operating constraint, the current process and the business result that would justify changing it.
Decide what teams should share
Read the shared AI infrastructure guide for the broader architecture, and reliable AI agents for completion and recovery checks. See company AI infrastructure implementation if you need these foundations built.
Inspect the implementation checks
Open the dated test record and read the selected original assertions. The check uses blank temporary databases and synthetic records. It covers persistence, stale edits, replay and transport access controls on the exported implementation, rather than a live client deployment.
Reproducible evaluation · October 29, 2026
How much context does one task need?
We tested the public implementation with one completed task and its evidence, then added unrelated projects. The query stayed the same: retrieve the Atlas workflow. Each project used synthetic task records and source text in a temporary database.
| Unrelated projects | Full response, bytes | Scoped response, bytes | Less returned data |
|---|---|---|---|
| 10 | 26,848 | 3,137 | 88.32% |
| 100 | 240,781 | 3,137 | 98.7% |
| 500 | 1,200,381 | 3,137 | 99.74% |
All three checks preserved the completed task and its supporting evidence, excluded unrelated tasks and left the underlying records unchanged. The scoped response explicitly said that it covered only part of the workspace.
The practical lesson
An agent resuming one project should receive the relevant work and evidence, including completed work that affects the decision. Sending every unrelated record makes the response grow with the whole workspace. Scope the read, then check what was preserved before treating a smaller response as an improvement.
Run the same check
Download the Python runner and inspect the dated results. The runner checks the implementation hashes before creating temporary databases. It needs Python’s standard library and the pinned public source.
git clone https://github.com/namankumar/agent-infrastructure cd agent-infrastructure git checkout 4cd0fb407a66ecdc788506f348ad7aad44d3ba33 python3 /path/to/scoped-retrieval-benchmark.txt --source .
Measured with Python 3.9.6, counting UTF-8 bytes of each JSON structured response. There was one query and one run per fixture size. These are payload measurements. Token use, response time, operating cost and task quality require their own measurements. The implementation still constructs an internal full snapshot before filtering; this test does not establish less database work or permission enforcement.