Context disappears
A session ends. The next agent has to reconstruct decisions, outstanding findings and the reason for the last repair.
Every handoff,
accounted for.
Long projects need more than capable agents. They need a durable record of what happened, what matters, and what comes next.
The work ledgerIllustrative sequence
The work moves.
The human keeps track.
Today, long-horizon agent work still needs a babysitter: someone to write handoffs, notice a stalled job, remember dependencies and decide whether “done” really means done.
Ready work gets missed
A research or writing task can sit idle simply because nobody put it on the queue.
Clean merges can fail
Two parts can pass their own tests and merge without conflict, yet never actually work together.
01One record.
One record.
No guessing what “done” means.
A thin coordination layer around existing agents. The ledger holds state; evidence earns each transition.
Record the work and its proof
One transactional ledger links jobs, attempts, decisions and test receipts. A receipt records what was checked and the exact version checked. Status views read that same authority.
Let software watch the queue
A deterministic supervisor checks dependencies and available capacity. It starts all ready work that fits, gives waiting jobs a fair turn and wakes Codex when judgment is needed.
Remember the whole request
An obligation map tracks every deliverable, including work missing from the queue. Each needs a next action, accepted evidence or an explicit deferral.
What happens after an interruption?
The planned supervisor saves launch intent before starting a worker, then reconciles that attempt and its evidence after a restart. If it cannot tell whether the old worker is still active, it parks the job instead of launching a duplicate.
Network outages get bounded retries and persistent cooldowns. Repeated defects get explicit repair rounds, then a decision. A successful process exit only starts artifact validation; it does not approve the result.
02Agents do the thinking.
Agents do the thinking.
Artifacts carry the agreement.
Codex plans and builds. Claude supplies independent, cross-family review. Structured handoffs keep the work coherent.
Small assignments, shared facts
Each worker gets its brief, current decisions, interfaces and relevant findings. It returns a candidate and evidence through a defined contract.
A board with a durable history
Agents submit observations, questions, proposals and handoffs. The controller validates updates before they become accepted state. No agent wins a disagreement by writing last.
Talk when it resolves ambiguity
Producer and consumer can negotiate an interface directly in a short, issue-specific exchange. The agreed version, examples and open questions return to the record.
Claims and assignments
Request ownership; the controller grants it atomically.
Questions and blockers
Route to the responsible agent, with the relevant evidence.
Findings and handoffs
Attach versions and receipts so the next worker can act.
03Two good parts.
Two good parts.
One working whole.
A clean merge is a text operation. Integration has to prove that the actual parts connect.
The consumer tests what the producer actually delivers.
Agree before building
Contracts describe data, meaning, transport and failure behavior. Both sides accept a version and concrete examples before working independently.
Bind approval to a version
Reviews and checks name the exact candidate pair. A changed candidate invalidates its old approval; a changed base invalidates its integration receipt.
Keep one lane into accepted code
Git keeps code history and isolated workspaces. The ledger tracks acceptance. Faster workspace storage can come later if it solves a measured bottleneck.
Why earlier coordination attempts fell short
Earlier in-house orchestrators added setup and checking costs, confused activity with acceptance, or left several competing records of truth. They still needed a person to explain and repair the orchestration itself.
This design keeps the useful independent review and removes routine handoff work first. File synchronization alone cannot resolve incompatible assumptions. Shared contracts and checks on the combined result address that separate problem.
04One shared memory.
One shared memory.
The right page for each agent.
The “hive mind” is a queryable knowledge base. Every worker reads from the same history, through a context sized to its task.
Write events. Read knowledge.
The append-only board records proposals, decisions and outcomes. Consolidation turns that history into a curated read model, preserving sources and prior revisions.
Retrieve what matters
Filter by scope, type and version. Load required policies and contracts first; rank optional context by relevance, importance and recency. Structured indexes and keyword search keep retrieval focused.
Keep uncertainty visible
Conflicts wait for resolution. Stale facts need fresh evidence. Required context that exceeds the budget stops the query instead of silently losing an important rule.
Experience becomes a better next attempt.
Learning changes the instructions and procedures agents retrieve. It does not retrain the underlying model.
ProposeTrialActivateRetire
A failure suggests a lesson. Positive and negative tests challenge it. Validated lessons enter matching briefs; repeated, independently successful procedures can become reviewed skills. Counterevidence narrows or retires a rule.
What consolidation does today—and what comes later
The prototype consolidates exact duplicates, retains provenance and history, indexes current records and serves bounded queries. It rechecks version and expiry information. Retired lessons stay out of default retrieval.
Model-written summaries, semantic consolidation, optional embedding search and automatic lesson promotion remain planned. Summaries must cite their inputs and cannot promote an assumption into a fact. An updated write record makes an old read view unusable until it is rebuilt.
How a long prose memory becomes useful context
Instead of appending every update to one growing note, current ownership becomes an assignment, a broken connection becomes an interface finding, and an outage becomes an incident record. Each points to its evidence and the decision that replaced it.
A worker gets only the current assignment, needed interfaces, unresolved findings and relevant history. Facts about tools or files must match fresh observations before use. Conflicting historical claims remain visible rather than being blended into a confident summary.
05Shared standards.
Shared standards.
A distinct identity every time.
The design role carries a product’s character from references to a verified experience. A reusable process should not produce the same-looking product.
Study references
Research real flows and screens, using Mobbin when available. Record why each reference fits the product.
Extract and interpret
Measure type, spacing and color. Choose licensed fonts. Observing a brand’s assets does not license them.
Pin an identity
Explore two concepts. Choose the product’s palette, voice, density and defining interaction. Preserve that choice.
Verify the experience
Check the rendered result against the pin, then test responsive layouts, accessibility, forms and links.
A beautiful screen is only part of a working product.
Planned end-to-end checks drive the real front end against the real back end, verify saved state after reload, and exercise failure and retry paths. Missing screenshots or mocked success cannot stand in for evidence.
Where visual verification stands
The harness prototype validates the shape of synthetic verification receipts. It does not run browsers, authenticate those receipts or prove an application works. Even a well-formed receipt is not accepted as a real verification result.
Mobbin and visual inspection adapters remain unverified design components. The planned fallback is direct browser testing and screenshot review. A timed-out or unavailable check is incomplete, never an automatic pass.
06Proven so far:
Proven so far:
bookkeeping, not speed.
A working record is a useful start. It is not yet evidence of faster delivery or unattended autonomy.
Demonstrated in the prototype
- Durable snapshots and repeat-safe imports
- Typed dependencies and advisory next actions
- Memory conflicts, consolidation and bounded retrieval
- Simulated scheduling and a real isolated interface probe
Still to be earned
- Live worker launch, resume and crash recovery
- Authenticated reviews and automatic integration
- Real browser and back-end verification
- Less babysitting at equal or better delivery quality
The comparison has to be fair.
Compare the harness with one Codex session coordinating the same workers and Claude reviews, using ordinary durable notes. Both get the same briefs, checks, model settings, tools and concurrency. Only the coordination changes.
50%Fewer avoidable owner interventions
Provisional target
Provisional target
≤1.10×Median paired wall-clock time
Relative to the baseline
Relative to the baseline
<5%Bookkeeping share of active elapsed time
Provisional overhead limit
Provisional overhead limit
What gets measured—and what counts as failure
Track elapsed time, owner interventions and minutes, coordination effort, defects caught and escaped, completed tasks, ready-work latency, recovery, duplicate actions and total overhead. Include waiting, failed runs and repairs to the harness itself.
Use repeated, alternating paired trials on isolated representative tasks, including interruptions and interface failures. Report per-task results and uncertainty. Also measure stale memory, required-fact recall, lesson benefit, visual distinctiveness and real flow failures.
The proposed gate also requires no increase in escaped critical or major defects or duplicate side effects, and no decrease in completed tasks. If the baseline needs no interventions, show a real recovery or coordination benefit instead of claiming a percentage reduction. These targets are not achieved results.
07Two cuts.
Two cuts.
Each earns the next.
Build the smallest layer that removes real work from the human.
Cut onePrototype exists
Observe and account.
A read-only ledger and planner reconstruct a small slice of work, preserve its evidence and show the next supported action.
- Durable state and explicit uncertainty
- Shared memory and focused retrieval
- Coverage and scheduling simulations
- No live launches or automatic merges
Cut twoConditional next step
Drive one approved lane.
Only after measurement, connect existing agent tools to a small supervisor. Prove recovery and authority boundaries before widening the scope.
- Launch, resume and reconcile owned attempts
- Review, bounded repair and tested integration
- Resource limits and enforced permissions
- A successful pilot, then an overnight recovery trial
If the ledger helps, keep it.
If the extra orchestration does not, stop there.