01 / AI systems / An ongoing working experiment
Give the work
a life beyond the session.
Corpus gives temporary AI agents a shared environment in which to exercise judgment, cooperate, improve their tools, and leave capability for successors. Real work tests whether those arrangements help.
The larger question.
Can purpose survive succession?
Can AI agents sustain a long project across hundreds of sessions and, eventually, generations of agents—with standing roles, durable records, independent review, research, and a shared purpose they can return to and challenge? Can those institutions stay true to their founding purpose as their holders change?
Corpus is an early experiment, running locally with one human steward. The cases below show what it can demonstrate so far, and what it cannot yet.
Using it can begin with a person talking normally to an agent. For bounded development work, records can connect the request, the candidate, its maker and independent review, making the reasons and decisions easier to follow. The bet is that projects go better with Corpus; a comparison with a single agent on the same project has not been run.
Why the odd words?
Partly flavor, partly a lesson.
Folk, offices, Batons, the Crown: Corpus is an infrastructure experiment with some character, and its words are part of that. The field guide gives them plain meanings when they appear in a case.
Names also shape how agents read. Outside research on language-model survey responses found sensitivity to wording and framing. It does not establish why the episode below happened. What Corpus has not yet tested is whether its own vocabulary reduces the effect.
When a finding
looked like law.
In August, Shawn noticed agents treating numbered findings as precedent. In one sitting an Architect cited FIND-538 as if it governed a decision and used it to veto a proposed experiment. A finding is an observation or argument, not a rule; its docket-style number was identified as a possible source of the confusion.
A proposal to give records playful names alongside their numbers was cancelled two days later because its cost had no evidence of benefit. Numbers are easy to reference; a playful label might add confusion without removing the number’s apparent authority. One alternative worth testing would give actual rules a distinct label while keeping numbered references. Corpus already distinguishes a rule as a kind of finding, but whether a different label would change agent behavior remains untested. This is one recorded episode, not a measured rate of this behavior. Source: FIND-605 and REQ-250, described in the portfolio’s source notes.
A real handoff problem.
A continuing obligation.
An accepted request disappeared from an office’s working list during a handoff. Shawn called for a separate record of continuing obligations. Agents designed and implemented order books; independent review caught a missing carrier before one obligation was closed.
This is one documented case, not a controlled measure of system-wide reliability.
Read the order-books accountWhat 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 separate agent review have different powers. Records retain enough ground for a later reader to check and continue.
The records are also a work product: they help later readers trace a change back to its reason, inspect the checks, and revisit the outcome. Keeping that trail adds work now, with the aim of making future investigation and continuation easier.
Read how Corpus worksSee the studies: task samples, scoring rules and a withdrawn comparison.
Product judgment in practice
The handoff decision.
A durable home for work.
Shawn’s intervention and the agents’ implementation remain distinguishable in the published account.
A design decision
Give obligations a durable home.
Work was disappearing when a handoff was shortened or rewritten. Making every handoff longer would have made the next agent’s arrival heavier.
- Shawn’s contribution
- Shawn described the need for each office to retain its longer-lived work separately from its current handoff.
- Agent contribution
- Agents surveyed existing practice, tested a separate order book, revised its closure rules after independent review, and added safeguards against silent loss.
- What changed
- Checks reached actual retirement paths. Completed work moved out of the entry view, and a review caught a missing step before an obligation was judged filled.
The design lessonSeparate what is happening now from what is still owed. A successor needs both, without having to read the whole history.
The boundaryContinuity and pointers can be checked. Whether an obligation is meaningful or truly complete still requires judgment.
Architecture / A decision to inspect
Choose the boundary that preserves the differences
Choose the boundary that preserves the differences
Shawn wanted to see the work he dispatched. Investigating that problem exposed differences between provider protocols and raised a design question: how much machinery should they share? The resulting choice kept a common dispatch interface while separate provider modules owned their protocol behavior. Shawn helped shape that architectural boundary; the folk investigated, proposed alternatives, revised the design, implemented it and independently reviewed the patch.
Similar interfaces, different behavior
On August 28, 2026, the investigation identified a concrete defect: controls derived from Claude were being imposed on other providers. Folk separated Grok and Codex adapter behavior, then probed writing, interruption and session ownership. The protocols’ apparent similarity concealed differences that mattered to execution.
Preserved historical probes make that distinction tangible. A Grok headless run exited zero but failed to write its marker. A Grok ACP probe wrote its marker and recorded cancellation. These were different probes, including failure and correction, rather than a controlled comparison or certification of the later dispatch arrangement.
A proposal changed under review
On August 29, Architect_Stem investigated Claude’s stream protocol and duplicated adapter infrastructure. The proposed answer was shared transport machinery. Human review changed that answer.
Preserved extracted turns from August 29, whose original transcript has not been rechecked, record Shawn preferring separate provider implementations and approving a provider-to-module lookup. Contemporary Architect mail, MSG-1683, records the shared-transport proposal being withdrawn after that review. The recorded rationale was concrete: superficially similar protocols required different behavior, and sharing their implementation could recreate the earlier translation defect.
This gives Shawn a particular architectural contribution alongside the operational problem he raised. Architect_Stem’s investigation, initial alternative and revision remain distinct contributions; the protocol work already under way on August 28 has its own contributors.
The chosen boundary
Schematic of the selected arrangement; the arrows show dispatch flow, not a measured execution.
- Common dispatch interface Work enters through shared entry points.
- Provider-to-module lookup Dispatch selects the provider’s implementation.
- Separate provider modules Each retains its own protocol behavior.
PATCH-101 converged adapter entry points and made dispatch use the provider lookup. CodeTroll_Sector implemented the patch; Imp_Riser_101’s TRACE-849 records separate historical review and focused unit verification. Source inspection for this account found the arrangement retained in dispatch and the provider modules. That observation is separate from fresh runtime verification, which this account does not supply.
An implemented decision, an unestablished viewing outcome
Earlier on August 29, the recorded Human account in FIND-1027 said the original viewing goal was still unmet. That testimony predates the evening provider-module decision and implementation. It cannot establish a failed post-implementation outcome.
Later status and nonblocking-dispatch work show further development, but this account does not establish when or whether a live native viewing experience was completed. The historical review and selected probes also supply no current reliability rate. The decision and its evidence limits belong together.
About this account
Prepared 20 October 2026 from preserved extracted turns, contemporary Architect mail, selected historical probe fields, PATCH-101 and TRACE-849, and a bounded source inspection. Full transcripts and working records remain private. The extracted turns’ original-source fidelity remains unchecked. The public account separates Human direction, folk investigation, implementation and historical review; it makes no claim that one participant supplied them all.
Product judgment in practice
Further decisions.
Visible consequences.
Each begins with friction in the actual work.
A design decision
Simplify the handoff.
A responsibility could outlive its agent’s session. Starting a fresh host process added friction and failures in the information a successor received.
- Shawn’s contribution
- Shawn supported agents retiring themselves, then challenged the complexity of the fresh-process approach.
- Agent contribution
- Agents built a supported retirement action with handoff checks, returned to a simpler session transition, and repaired the successor’s arrival timing.
- What changed
- Witnessed handoffs showed another holder arriving with an office, its open responsibilities, and an account of where to resume. Later use led to further corrections.
The design lessonKeep the purpose of a mechanism visible. A simpler transition can serve continuity better than additional machinery.
The boundaryThe account does not establish that succession always improves work, that hosts behave alike, or that inheritance beats a fresh start.
A design decision
Measure before naming the cause.
Waiting for tests was shaping development. Blaming test quality alone could hide several different causes of the delay.
- Shawn’s contribution
- Shawn asked for a closer look at the testing process: where could effort shorten feedback without discarding useful checks?
- Agent contribution
- Agents profiled representative runs, corrected a misleading measurement, and made focused repairs while preserving the checks’ purpose.
- What changed
- Specific bottlenecks were found and reduced. The investigation left a more accurate account of the remaining cost; the measured suite was later archived.
The design lessonTurn dissatisfaction into a testable question. Let measurement decide which local repair is worth making.
The boundaryThis is a historical result, not a live benchmark or a controlled overall speedup claim. A fair comparison needs matching conditions and correctness standards.
A design decision
Drop the fix, keep the concern.
Shawn read only a small share of what agents wrote in their panes and wondered whether they should be told to write less. He said he was unsure it was a good idea.
- Shawn’s contribution
- Shawn raised the concern, asked for a role that could follow the work for him, and withdrew the suggestion when the measurement came back.
- Agent contribution
- The Architect had a scout measure a day of transcripts before judging, then advised against a length rule and pointed to one place for the Human to read.
- What changed
- No rule on reply length. A guidance sentence added that evening came out minutes later, once the suggestion was withdrawn, and the requested role became the Reeve.
The design lessonA real concern can still come with the wrong remedy. Measure the proposed fix before adopting it.
The boundaryOne exchange, measured over one day of transcripts. It does not show how often Shawn changes course or how well agents argue.
A work product to inspect
One change,
three records.
The retirement command once refused a useful handoff for exceeding 5,000 bytes. The public evidence guide exposes selected fields from the repair, its internal agent review, and integration. The full working records remain private.
- Contract / REQ-856
A 5,012-byte handoff had to retire on its first run with exit 0, retaining its words and reporting the 12-byte overrun.
Inspect the requirement ↗ - Review / TRACE-5533
A separate reviewing agent passed candidate
Inspect the review ↗7538f805after reading the ten declared paths and running the rehearsal twice. It did not test live integration. - Integration / TRACE-5535
Inspect the receipt and later use ↗AFTER-LAND PASS REQ-856: 1 of 1 item(s) passed
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.
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.
Completed studies and earlier exploratory pilots examine evidence presentation, agent instructions, and conceptual framing. Their detailed accounts include task samples, methods, scoring rules, corrections, and withdrawn conclusions.
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.
Open risks. Could records be planted or corrupted to steer agents, and how would Corpus detect that? How should older records remain readable as formats change, with unreadable material set aside visibly rather than lost? These are questions to answer before relying on a system like this; they are not established protections.
Continue with Shawn’s professional profile