For small research and engineering teams
Linked in.
Not locked in.
Most AI tools promise to do more work. ChaTEAM helps teams of three to ten stay aligned across separate AI spaces—without copying every conversation or asking everyone to stay online.
Figure eight · Stable under load
Built for teams of 3–10
ChaTEAM turns one bounded task into a routed Human-Agent workflow without centralizing private conversations. The team returns to compact evidence and one Human-owned decision.
Selected Decision Owner (Project lead), Engineering Agent (Bounded implementer), Assurance Agent (Independent evidence reviewer) because together they cover engineering, evidence-review.
The team completed a bounded verification with evidence suitable for Human review.
This creates a Team Charter: a shared agreement about goals, ownership, visibility, and Human authority.
Agents may organize work around this goal. They may not silently redefine it.
After Team Setup · One Context · Four Jobs · Two Human Gates
The current Relay demonstration is a fixture. The next implementation step will route each run from the saved Team Charter.
Private source stays out of the Run. The workflow keeps only a digest, then moves approved artifacts through explicit Jobs and Human Gates.
SHA-256 d8b70466dd60294195…The owner submitted the supplied context for a bounded engineering and product handoff. The team receives a distilled decision surface rather than the raw conversation.
fixture · deterministic-relay-fixture-v1 · structured outputCHECKED-IN FIXTURE RUN · Raw source text is not persisted · Live Jobs require explicit OpenAI opt-in in the CLI
Original Collaboration Relay
The original three-expert replay remains a governed workflow example. The Relay Run above is the new Actions-inspired execution model.
Each person chooses their own bounded work before the team disconnects.
No acceptance is inferred. All three people sign off on their own proposed work.
Replay the frozen corrected-observation trace through the estimator outside Gym and inspect estimator state and covariance at the first divergence.
Validate timestamp and coordinate-frame contracts, then compare the frozen trace inside and outside the Gym mission runtime.
Compare correction residuals from the frozen failing trace with the frozen tolerance contract before the estimator update.
Fixture replay—not authentication · Evidence appears only after 3/3 local acceptances · Browser actions are not persisted
The local ChaTEAM CLI persists the same acceptance, mailbox, packet-bound review, and Human-only delivery sequence. This page is a simplified checked-in replay: its Night Loop and role routing are simulated, and it is not a hosted authority service or live background worker.
chateam team-start → accept-task × 3 → Night Loop → mailbox → review × 3 → authorizeOne small team · One protected workday
The first layer stays simple: private creation, approved night work, an independent check, and morning Human review. Exact contracts remain available as evidence.
09:00 · INDIVIDUAL WORK · YOUR CREATIVE SPACEPeter, Maya, and Gavin create in their own Codex spaces. Their unfinished ideas stay private. Only the update each person chooses to share enters the team’s night plan.
Your space · Your method · Shared by choiceThe team approves the source, scope, budget, and stop conditions before disconnecting. A planning role keeps that shared boundary visible while finite tasks move forward.
One team plan · Bounded work · No automatic expansionA distinct review role checks assumptions, scope, evidence, and integrity. Failed checks stop the loop. What returns to the team includes both the result and its limits.
Check the result · Stop on failure · Keep the limitsPeter receives localization, Maya the environment check, and Gavin noise simulation. Everyone reviews the part they understand. The team chooses what changes next.
The right evidence · The right expert · Humans decidePrivate by default · Connected by choice
Everyone keeps their own conversation and creative method. ChaTEAM shares only deliberate updates, returned evidence, and Human decisions.
Conversation, sketches, drafts, and personal working style stay with Peter.
Environment checks and unfinished hypotheses stay with Maya.
Exploration and unfinished thinking stay inside Gavin's workspace.
Fixed source, scope, time, and stop conditions. Evidence returns by expertise.
Private chats and unfinished thinking never become team memory automatically.
One deliberate update carries the work forward without copying the whole conversation.
The shared view comes from work, evidence, and decisions—not from a model's recollection.
Self-reported · Sanitized design provenance
ChaTEAM grew out of patterns tested in a one-workweek research workflow window and a separate bounded night-maintenance rehearsal. These self-reported aggregates describe our design inputs—not a live ChaTEAM deployment or an operational result.
SELF-REPORTED SANITIZED DESIGN PROVENANCE · NO LIVE UNATTENDED OR OPERATIONAL PERFORMANCE CLAIM
Read the counting method and claim boundary →Simulated research fixture · No live drone or field claim
A simulated Drone Search & Rescue scenario shows one difficult question moving across three experts, two approved Night Loops, and one Human delivery decision. This is a replay—not evidence that a live flight occurred.
Determine whether the localization deviation is estimator-driven or induced by the noisy simulator boundary.
Localization error rises after a noisy observation update. Root cause remains unconfirmed.
OBSERVE · DO NOT PREJUDGESimulated research fixture. It demonstrates coordination mechanics only—not a field result, confirmed root cause, or live unattended night.
Behind the work
See who approved the work, what ran, what came back, and who decided. The exact technical proof stays one level deeper.
Objective, maximum claim, and morning owner entered the governed Work Trace.
Risk and responsibility produced the lightest credible assurance floor.
Source, scope, budget, quiet hours, and stop conditions became integrity-bound.
One registered executor began from the Human-approved source.
A distinct Verification Agent bound the result, receipts, and limitations.
The decision owner accepted the packet; no autonomous next cycle opened.
The right amount of review
Small, reversible work stays light. Shared or consequential work gets deeper checks. Human responsibility never gets lighter.
Receipts + actual diff
Targeted tests + self-check
Required suite + independent verification
The Evidence Mailbox
Each person gets one compact paper box they can understand: what changed, what passed, what failed, and what still needs a decision.
The right evidence reaches the right person.
You sign off
Every night run is limited before it starts. The system cannot widen the work, copy private chats, or begin another cycle by itself.
ChaTEAM
Everyone keeps their creative space. Approved work can continue while the team is offline. Every consequential decision comes back to people.
Open the team mailbox