01 / AI systems / An ongoing working experiment
Give the work
a life beyond the session.
Corpus connects temporary AI agents to durable responsibilities, shared tools, and evidence that another agent can recover.
What happens when
the agent leaves?
Useful agent work depends on more than the answer at the end of a turn. There is a contract to meet, a candidate to inspect, a decision to preserve, and often unfinished work to continue. A session can end while all of those remain.
Shawn originated Corpus and shaped the environment with the agents working inside it. Offices carry durable functions; temporary holders take them for a tenure. Accepted work has a scope. Making and independent certification are different powers. Records retain enough ground for a later reader to check and continue.
See how Corpus worksAn interactive reading / three documented cases
Choose the intervention.
Follow what it changed.
Each path separates Shawn’s product choice, the agents’ work, the recorded outcome, and the limit of the evidence. Select a case and a step to examine the link.
Case 01 / Order books
Give obligations a durable home.
A handoff describes the current situation. When it is rewritten, a longer obligation can quietly disappear. The house’s survey found work dropping from handoffs and during sittings.
01 / Human choice
The intervention
Shawn proposed that every office keep the work it owed separately from its current handoff. He also raised review as the basis for cancelling an obligation.
02 / Agent response
The work
Agents surveyed existing records, designed the order books and history checker, corrected the first draft through independent readings, and connected the check to entry, retirement, and landing.
03 / Observed outcome
What the record shows
The checker flags an open order removed without a recorded close or named destination. Closed work can be shelved while its history stays recoverable.
04 / Limit
What remains open
The checker reads identifiers and pointers. It can catch a dropped order; it cannot establish that a claim of completion is semantically true.
Read the dated account: Order books . The path and stage labels are portfolio interpretation; the account carries its evidence and unresolved edges.
Case 02 / Succession and retirement
Renew the session. Keep the process.
Replacing a long-running holder required a reliable handoff and a new session. A fresh-process approach introduced missing office cards, stale pane names, and quit-dialog friction.
01 / Human choice
The intervention
Shawn challenged the restart approach and chose session clearing with the process retained. His intervention redirected a mechanism that had become more costly than its purpose warranted.
02 / Agent response
The work
Agents built the retirement commands, preserved handoff and work records, implemented the delayed successor wake, removed the fresh-launch path, and repaired the edges found in use.
03 / Observed outcome
What the record shows
The recorded succession route cleared the session, then waited to deliver the next holder’s entry. The office and its current ground continued through a change of holder.
04 / Limit
What remains open
The published account is dated and names unresolved host differences and later changes. A working route on one host does not prove an equivalent contract on every provider.
Read the dated account: Succession and retirement . The path and stage labels are portfolio interpretation; the account carries its evidence and unresolved edges.
Case 03 / Tests under the floor
Measure before naming the cause.
Slow tests interrupted the development loop. A large suite made it easy to accept delay as inevitable or blame test quality without identifying the actual cost.
01 / Human choice
The intervention
Shawn raised an explicitly unevidenced hunch about the testing process and pushed for less time spent waiting. He supplied the problem and the ambition; measurement had to establish the explanation.
02 / Agent response
The work
Agents profiled setup and execution, inspected repeated database initialization and fixture scheduling, tested explanations one variable at a time, repaired defects, and preserved disproven readings in the account.
03 / Observed outcome
What the record shows
The investigation exposed concrete costs and regressions, including repeated store setup and a dominant fixture. It also corrected an assumed worker count and separated performance work from correctness failures.
04 / Limit
What remains open
The runs differed in commits, configuration, and host conditions. This portfolio makes no controlled overall speedup claim and does not claim the aspirational latency target was met.
Read the dated account: Tests under the floor . The path and stage labels are portfolio interpretation; the account carries its evidence and unresolved edges.
Product judgment in practice
Three decisions.
Visible consequences.
Each begins with friction in the actual work. Shawn’s intervention and the agents’ implementation remain distinguishable in the published account.
A design decision
Give obligations a durable home.
A handoff describes the current situation. When it is rewritten, a longer obligation can quietly disappear. The house’s survey found work dropping from handoffs and during sittings.
- Shawn’s contribution
- Shawn proposed that every office keep the work it owed separately from its current handoff. He also raised review as the basis for cancelling an obligation.
- Agent contribution
- Agents surveyed existing records, designed the order books and history checker, corrected the first draft through independent readings, and connected the check to entry, retirement, and landing.
- What changed
- The checker flags an open order removed without a recorded close or named destination. Closed work can be shelved while its history stays recoverable.
The design lessonA memory system needs places for different kinds of memory: current state, continuing obligations, and historical accounts. Making those distinctions changes what the software can protect.
The boundaryThe checker reads identifiers and pointers. It can catch a dropped order; it cannot establish that a claim of completion is semantically true.
A design decision
Renew the session. Keep the process.
Replacing a long-running holder required a reliable handoff and a new session. A fresh-process approach introduced missing office cards, stale pane names, and quit-dialog friction.
- Shawn’s contribution
- Shawn challenged the restart approach and chose session clearing with the process retained. His intervention redirected a mechanism that had become more costly than its purpose warranted.
- Agent contribution
- Agents built the retirement commands, preserved handoff and work records, implemented the delayed successor wake, removed the fresh-launch path, and repaired the edges found in use.
- What changed
- The recorded succession route cleared the session, then waited to deliver the next holder’s entry. The office and its current ground continued through a change of holder.
The design lessonA product decision can be a deletion. Start with the user-visible purpose, then ask which machinery is earning its complexity.
The boundaryThe published account is dated and names unresolved host differences and later changes. A working route on one host does not prove an equivalent contract on every provider.
A design decision
Measure before naming the cause.
Slow tests interrupted the development loop. A large suite made it easy to accept delay as inevitable or blame test quality without identifying the actual cost.
- Shawn’s contribution
- Shawn raised an explicitly unevidenced hunch about the testing process and pushed for less time spent waiting. He supplied the problem and the ambition; measurement had to establish the explanation.
- Agent contribution
- Agents profiled setup and execution, inspected repeated database initialization and fixture scheduling, tested explanations one variable at a time, repaired defects, and preserved disproven readings in the account.
- What changed
- The investigation exposed concrete costs and regressions, including repeated store setup and a dominant fixture. It also corrected an assumed worker count and separated performance work from correctness failures.
The design lessonA useful demand does not have to arrive with the right diagnosis. The system should convert dissatisfaction into a testable question and let evidence change the answer.
The boundaryThe runs differed in commits, configuration, and host conditions. This portfolio makes no controlled overall speedup claim and does not claim the aspirational latency target was met.
The difficult part
A persistent system
can preserve its mistakes.
The hard design problem is deciding what should endure, who can change it, and how a later reader can tell whether it remains useful.
Order books made dropped obligations detectable, then accumulated completed material at entry. Succession machinery gained complexity, then needed a simpler route. Performance explanations sounded plausible, then changed under measurement. The same persistence that supports continuation can carry stale assumptions and unnecessary work.
The response in these cases was to revisit the mechanism, keep the correction visible, and make the boundary smaller or clearer. That is a practical design discipline: a preserved story must remain answerable to what actually happens.
Give responsibility a home. Make the candidate identifiable. Let review change the answer. Keep enough history to understand the correction.
Why this matters to AI
Design the environment
around the capability.
More capable models still work through interfaces, assignments, permissions, source state, and human decisions. Corpus makes those surroundings a product concern. It provides concrete material for studying handoffs, responsibility, recovery, and the quality of agent-produced work.
Shawn’s contribution is visible in choosing those problems and challenging mechanisms in use. The agents provide another kind of evidence: they implement the system, encounter its friction, and leave records that can be compared with its intended behavior.
What comes next / Portfolio perspective
Test the structure
against simpler practice.
Corpus demonstrates that this environment can be built and used. These cases do not establish that its full structure outperforms a simpler human–agent loop, or that inheritance improves every task.
A useful next step is a fair comparison on work beyond Corpus: similar tasks, models, rounds, and Human attention, with independent judgment of defects, rework, cost, and continuation.
- Which responsibilities need a durable owner?
- What can be removed without losing a necessary check?
- Does the next holder recover useful ground at a reasonable cost?
- When does the structure improve a result enough to justify itself?