Skip to content

My First 13-Hour TeaParty

I just ran TeaParty's first end-to-end test. The goal was to see whether the agents could take a real prompt from intent through plan and execute to a coherent deliverable. The prompt:

I would like a book on the universal nature of humor. There are certain types of humor that transcend time, culture and language (e.g. comedy wildlife photos, physical humor, affiliative humor, etc). This is a 5-7 chapter book, targeting armchair enthusiasts, and should explore this phenomena across cultural, temporal, language, belief, and technological boundaries. Thesis: Humor is what unites us.

What TeaParty is

TeaParty is the multi-agent orchestrator I've been building. It runs on the Conversation for Action protocol: three phases (intent, plan, execute), each gated by approval. A project lead breaks the work into tasks and dispatches sub-agents into isolated worktrees, where each runs its own tactical plan-and-execute (no nested intent phase). The proxy stands in for me at task-assert and most other escalations. Work-assert is mine.

What I observed

Below is a gantt chart of the run. Intent gathering took 5 minutes; planning took 16; execute ran for just under 13 hours before closing at the work-assert gate. Total runtime came in at 13 hours 6 minutes. Of that, 7 hours were spent at the work-assert gate waiting for me to wake up.

gantt
    title 13h run — Sun 17:10 PDT to Mon 06:16 PDT — one row per agent
    dateFormat HH:mm
    axisFormat %Hh

    section Orchestration
    project lead                              :active, 00:00, 786m

    section Intent + Plan
    intent agent                              :done, 00:00, 5m
    plan agent                                :done, 00:05, 16m

    section Research lead
    research lead                             :active, 00:22, 244m

    section Research wave 1 (8 parallel, all failed)
    prologue brief W1                         :crit, 00:22, 35m
    ch1 brief W1                              :crit, 00:22, 35m
    ch2 brief W1                              :crit, 00:22, 35m
    ch3 brief W1                              :crit, 00:22, 35m
    ch4 brief W1                              :crit, 00:22, 35m
    ch5 brief W1                              :crit, 00:22, 35m
    ch6 brief W1                              :crit, 00:22, 35m
    ch7 brief W1                              :crit, 00:22, 35m

    section Research wave 2 (8 parallel, 3 ok)
    prologue brief W2 fail                    :crit, 00:57, 58m
    ch1 brief W2 fail                         :crit, 00:57, 58m
    ch2 brief W2 fail                         :crit, 00:57, 58m
    ch3 brief W2 ok                           :done, 00:57, 30m
    ch4 brief W2 fail                         :crit, 00:57, 58m
    ch5 brief W2 ok                           :done, 00:57, 30m
    ch6 brief W2 ok                           :done, 00:57, 30m
    ch7 brief W2 fail                         :crit, 00:57, 58m

    section Research wave 3 (5 parallel, 4 ok)
    prologue brief W3 fail                    :crit, 01:55, 123m
    ch1 brief W3 ok                           :done, 01:55, 60m
    ch2 brief W3 ok                           :done, 01:55, 60m
    ch4 brief W3 ok                           :done, 01:55, 60m
    ch7 brief W3 ok                           :done, 01:55, 60m

    section Research late retries (prologue only)
    prologue brief W4 fail                    :crit, 03:58, 17m
    prologue brief W5 ok                      :done, 04:15, 11m

    section Writing lead
    writing lead                              :active, 03:59, 547m

    section Spec writers
    umbrella seven-chapter spec               :done, 03:59, 30m
    prologue spec                             :done, 04:26, 10m
    ch1 spec                                  :done, 04:27, 10m
    ch2 spec                                  :done, 04:27, 10m
    ch3 spec                                  :done, 04:27, 10m
    ch4 spec                                  :done, 04:27, 10m
    ch6 spec                                  :done, 04:27, 10m
    ch7 spec                                  :done, 04:27, 10m
    ch5 spec (delayed 11 min)                 :done, 04:38, 10m

    section Draft writers (Phase 3 — starts after all specs incl. ch5 land)
    draft prologue                            :active, 04:48, 17m
    draft ch1                                 :active, 04:48, 17m
    draft ch2                                 :active, 04:48, 17m
    draft ch3                                 :active, 04:48, 17m
    draft ch4                                 :active, 04:48, 17m
    draft ch5                                 :active, 04:48, 17m
    draft ch6                                 :active, 04:48, 17m
    draft ch7 (post-snapshot)                 :active, 04:48, 17m

    section Editorial reviewers (Phase 4, parallel — read full manuscript)
    editorial 1 Senior Editor                 :done, 05:05, 6m
    editorial 2 Sunday report                 :done, 05:05, 6m

    section Revision agents (Phase 4 single revision pass)
    revise ch1 to _revised                    :done, 05:11, 4m
    revise ch2 to _revised                    :done, 05:11, 4m
    revise ch4 to _revised                    :done, 05:11, 4m
    revise ch5 to _revised                    :done, 05:11, 4m
    revise ch7 to _revised                    :done, 05:11, 4m

    section Verifiers (Phase 5 — sequential, both read _revised, both finish before gate 1)
    verify 1 → verification_report.md (FAIL on 3 sub-criteria) :crit, 05:15, 3m
    verify 2 → final_report.md (PASS on full INTENT + invariants) :done, 05:18, 7m

    section Gate cycle (deflection loop)
    gate 1 → ch7 Camus + Abol Tabol fix       :active, 05:25, 5m
    gate 2 → ch4_final + ch5_final + ch7_final :active, 05:30, 5m
    gate 3 → ch4 Gottfried rewrite            :active, 05:35, 4m
    gate 4 → gates 5 + 6 confirmation         :active, 05:39, 7m
    gate 7 task-assert approve                :milestone, 05:46, 0

    section Skill extraction
    skill candidate generated                 :milestone, 06:09, 0

    section Idle (human asleep)
    agents waiting at work-assert gate        :crit, 06:09, 417m

    section Final
    work-assert approve (human returns)       :milestone, 13:06, 0

Hierarchical dispatch with workspace isolation

Hierarchical dispatch is how TeaParty controls context. Each level sees only what it needs: the project lead works at strategy and orchestration; domain leads scope tasks for their slice; workers see only their own task. No sibling-to-sibling pollution, no parent-to-worker context bleed. At peak the dispatch tree reached three levels (project lead → domain lead → workers), each agent in its own isolated workspace, outputs merged back at gate boundaries.

The dispatch layer's fan-out was too wide. Launching eight concurrent research workers at once hit provider rate limits, and all eight failed in wave 1. Over four more waves the briefs came in as windows reset, with the prologue not completing until wave 5 about four hours after the first attempt. The system finished the work without a designed recovery path.

A subtler issue appeared during fan-in of the draft documents. Six of seven chapter specs completed inside a fifty-second window; the seventh was eleven minutes late. The fan-in was a polling loop that blocked until every child responded. Six chapter drafts couldn't start until the seventh spec arrived, even though their own inputs were ready. The loop also burned CPU for the full eleven-minute wait.

Orchestration of all three CfA phases

The three phases ran in order, with INTENT.md and PLAN.md on disk for the verifiers to audit. But agents burned tokens looking for files that had been written to unexpected locations. The Ch7 research agent, hunting for sibling chapter briefs in the expected research directory, didn't find them and noted as much in its own output: "Prior chapter briefs not located in accessible research directory; synthesis hooks draw from chapter descriptions provided in project brief." The verifier looking for INTENT.md at the worktree root flagged it explicitly: "INTENT.md does not exist at the worktree root. It is located at the main project root." Both agents recovered, but only after wasting tokens and working with incomplete context. Document layout belongs in the plan.

Learning, both procedural and preferential

This was the first end-to-end run, so there were no cross-run learnings to retrieve. However, the learning system fired as expected, storing memories in three different layers.

The first was three alignment-observation entries appended to the procedural-memory file (dated 2026-03-16), framed as feedback for future intents. One: "briefs will be attempted in parallel, with automatic retry on infrastructure failures" — the system correcting its own model of how parallel dispatch behaves. Another: "Intent stated 'I'll verify coverage' before Phase 2 — no coverage verification checkpoint was implemented or observed." The system noticing where the intent it received and the work it performed had quietly diverged, and writing the gap down for the next intent gate to see.

The second was six differentials in proxy-confidence.json across the PROPOSAL (2) and DRAFT (4) states. Each entry logs what the proxy predicted I would say, what I actually said, and the delta. From one of the intent-clarification questions:

Proxy predicted: "Option C — the hybrid… the types should be the backbone of the middle chapters."

Actual response: "I think organized by type works."

Outcome: clarify (proxy escalated rather than auto-approving).

"prediction_correct_count": 0 over "prediction_total_count": 2 is exactly the signal you want: specific predictions, told they were wrong, ground truth on disk for the next run.

The third was a skill candidate generated at end-of-run, serializing this run's PLAN.md as a pattern keyed to the original prompt. The extractor stored this learning before I'd approved the work. Had I rejected at the gate, the skill would already have been in the catalog and would shape the next run from a base of unapproved work.

Human proxy handling escalations

The proxy picked up nearly every escalation on my behalf: sub-agent questions through the run, plus the gate decisions that would otherwise have woken me. Where it got through, the decisions it raised matched what I would have raised at the same gate.

The strangest thing I watched happened at the task-assert gate during the revision pass, and it was the proxy failing to pick up. Six task-assert iterations in twenty-one minutes. The first was clean. The proxy picked up the agent's escalation, reviewed the verifier output, and raised the issues the writing team needed to address. Then the proxy stopped getting through. The agent hit a timeout during the next gate interaction, recovered, and saw an empty HUMAN: entry. The proxy's response had been lost to the timeout, leaving a blank where the response should have been. From the agent's side, no one was picking up. It thought I'd hit enter without typing anything, and answered:

"I'm not sure I can answer that right now. Could you rephrase, or let me know your decision?"

It said this six times. The phantom apologies stacked in the agent's input. First one pair, then three, then three again, because each gate iteration recovered from another timeout and saw another empty HUMAN: entry. Meanwhile, in its own trace, the proxy was reasoning about what was happening. By the sixth gate it wrote, in the middle of an otherwise substantive review:

"The human has entered three empty/null messages and gotten non-answers each time. The human is likely frustrated at this point and would want the gate to just work."

The proxy was modeling someone who wasn't there. By the fifth and sixth iterations its responses were substantive: citing line numbers, comparing files, confirming named changes. The agent never saw any of them; the channel never recovered. The work the proxy was asking for was happening between every pair of gates regardless. The loop was a phantom on top of real work.

The pattern generalizes. Any async dialog protocol where silence-from-human is indistinguishable from silence-from-timeout will produce both pathologies: phantom interlocutors on the agent side, modeled non-presence on the proxy side.

Active work concluded around 23:20 PDT Sunday. Task-assert had cleared at 22:56; the skill candidate auto-extracted at 23:18. After that the system kept running, waiting at the work-assert gate. That gate is mine, not the proxy's, and I was asleep. Six hours and fifty-eight minutes later, at 06:16 PDT Monday morning, I came back and signed off.

What it produced

I read the whole 55,000 words. Prologue plus seven chapters: Knock Knock, Born Laughing, The Oldest Joke in the World, Banana Peels and Power, You Had to Be There, The Last Laugh, Silence Is Funny, Spam Spam Spam.

The thesis surprised me. I had asked for humor as universal across cultural and temporal boundaries. The manuscript reframes around the two-beat structure of setup and punchline as the spine of every chapter, and the reframe worked. I had not seen it coming.

The voice did not. On the page it reads warm and engaging, but the warmth has no one underneath it. The "personal anecdotes" are not mine. I never told the agents about my life, but they wrote in first person anyway. The chapters lack a narrative throughline; examples and transitions follow no consistent author logic.

The gap is me. I ran the whole thing without showing up. Stylistic feedback could have come at task-assert. Anecdotes at the plan gate. A backtrack from EXECUTE to PLAN when the voice faltered. Three places to step in, and I stepped into none of them.

This run tested what the system does when the author walks away. It produced a manuscript with no author in it.

Lessons learned

The skill extractor fired before human approval. At 23:18 PDT a candidate was written to disk, almost seven hours before I cleared the work-assert gate Monday morning. Had I rejected, that candidate would already be in the catalog, contaminating future runs. Skill extraction must run downstream of human approval, never on auto-trigger.

Fan-in needs to be non-blocking and low-CPU. The polling loop in this run cost both wall-clock and compute: six drafts waited eleven minutes for one slow spec when their own inputs were already ready, and the loop burned cycles for the full wait. A message bus would fix both. With one in place, parents sleep until children respond, and each child's work proceeds the moment its dependencies arrive.

Intent and planning must specify where files live. When the plan doesn't say where artifacts go, agents make different assumptions. One sub-agent couldn't find sibling artifacts it needed and worked from less context than it should have. Others wrote the same kind of file under different names, leaving the filesystem with parallel versions of what should have been single artifacts. The confusion and the duplication share one root cause: an unspecified path. The fix belongs in INTENT.md and PLAN.md, not in the orchestrator.

Source

Full session record (state-machine trace, dialog transcripts, manuscript, editorial and verification reports, obstacles, and learnings) is in the TeaParty case-study v1 walkthrough.

Comments & reactions