← Back to Can I trust what the AI suggests and builds?
How do I read what an agent hands back?
Short answer
In an Agent.Coder or Agent.Concept reply, read what is done and not done with each tag, the standards followed, its Captures, and its last line.
Before you start
A reply from an agent in GilbApp's chat, after you asked it for work, as the answer on chatting with an agent in GilbApp shows.
Words used here (8)
- agent
- An agent is your AI following one written procedure, its standing instructions, for one kind of work.
- GilbApp
- GilbApp is the app, opened in a web browser at https://gilb.com/app, where you open, fill and share your vaults.
- Jackfruit
- Jackfruit is the vault agent: it finds, reads, writes and links records in your vaults, and makes new vaults for you.
- short name
- A short name is the one word you give one AI on one computer, such as gilb-example; it keeps that AI's key and tools apart.
- Kai-Zen
- Kai-Zen is the planning agent: it finds out what you want, chooses the way, hands the work to other agents and reports back.
- project
- A project is work to bring about a change you want, such as a new app or a better way of doing something.
- session
- A session is one stretch of work with an agent, from its start until you close it; with Kai-Zen it ends in a handoff.
- handoff
- A handoff is the note Kai-Zen writes when a session ends: what was done, what is open, and where to start next time.
Steps
Read what is done and what is not, and each tag.
In the agent's reply in the left pane, read what it says is done, what is not, and why. Each done carries REAL if it ran on the real thing, MOCK if it ran on a stand-in, or UNPROVEN if nothing was run.
More about this
An agent is your AI following one written procedure, its standing instructions, for one kind of work. For each file it saved, it gives the file's path on your computer or its address in a vault. Agent.Coder, which builds apps, and Agent.Concept, the agent that finds and makes GilbVeda's terms (GilbVeda is the shared list of terms), hand back this way; you can choose both in the chat. Kai-Zen's own report also says what is done, what is not and why, and tags each done; Jackfruit's is different (Deeper). GilbApp shows the reply as plain text: an address in it is not a link, but you can select it and copy it.
Read the standards it followed, and any it set aside.
Read which standards the agent says it followed, each by its tag, its short name such as Kai-Zen.EvidenceRules, and where its work shows each one. If it set a standard aside, it names the person who allowed that; an agent cannot allow it.
More about this
Agent.Coder and Agent.Concept list these; Kai-Zen and Jackfruit do not. A standard is a written procedure, rule set, form or other guide that people and agents follow as they work. A gap or a conflict the agent found in a standard is written down as a finding, a note kept with the project saying where, what it chose and why.
Read the Captures at the end of the reply.
Last in the reply, under Captures, the agent lists what it wrote down for the next session, such as an address, a tool on your side, a preference or a trap. It keeps each in the project's own folder, a vault or a folder on your computer, with where it came from and its date.
More about this
Kai-Zen's and Jackfruit's replies have no Captures. It never writes a secret there, such as a password or a key. What a step of the project taught goes in that step's Learned line instead, where Kai-Zen writes what worked and what did not. The answer on ending a session shows the handoff Kai-Zen writes for the next session.
Check that the last line names the agent.
The last line of the reply is the name of the standard the agent loaded, such as Kai-Zen or Jackfruit, alone on that line. A last line [No agent — base assistant] means it loaded no standard.
More about this
An agent signs only a name it loaded; an agent you made yourself signs only if its own steps say so. When the agent answers without all its standing instructions, GilbApp says so above the message box; the answer on seeing which standing instructions an agent loaded explains it. When your own AI answered, a small line under the reply names its model and ends with your Claude subscription or your SuperGrok.
Deeper
Also asked:
What does my AI's report at the end of a step mean? It says what is done, what is not, and why. Each done is tagged REAL (ran on the real thing), MOCK (ran on a stand-in) or UNPROVEN (nothing was run), and the last line names the standard the agent loaded.
Can I see a list of everything an agent did in my vault? No: GilbApp keeps no such list. Read the agent's report in the chat, kept in this tab for its last 32 messages; Jackfruit's lists every record it wrote, linked or ended.
A handback is what an agent says when it gives finished work back to whoever asked for it: you, or another agent. Jackfruit's standard asks for no tags: its report lists every vault it made, every access it gave or ended, and every record it wrote, linked or ended, with its address, and says what it did not do and why. When Kai-Zen hands work to another agent, that agent hands back to Kai-Zen, and Kai-Zen takes the work back, hands out the next work at once, and reports to you when the run, your request carried to its end, is over. In GilbApp's chat you see only the reply of the agent you chose. Only an agent whose standard tells it to hand back by Kai-Zen.Return lists the standards it followed and its Captures: Agent.Coder, Agent.Concept and Agent.Tester, the agent that tests what others built, do; the standards of Kai-Zen and Jackfruit do not ask for either, but both sign the last line the same way.
When Agent.Coder hands back a piece of an app it built, it also says how it was tested: a PASS from Agent.Tester, or self-QA when it had to test the piece itself, which is not equal to a PASS from Agent.Tester. Agent.Tester always gives its verdict together with the path or address of its saved report. For a change to an app's screen, Agent.Coder adds the path of a screenshot and the count of errors the browser reported. Both name any standard they could not read.
GilbApp keeps the chat in this browser tab only, and only its last 32 messages and replies. What lasts is what the agent saved, at the paths and addresses it lists. The answer on what counts as proof that something works explains the three tags.
What Kai-Zen itself sends you while it works, and what to answer: See: Kai-Zen says something I do not understand, or stopped. What now?
This answer is written for a computer screen.
Source: Standards read 2026-11-10: Kai-Zen.Return (K 06gj2sxn41x8v45wk68s4v7t2c) steps Report ("Say what is done, what is not, and why. Tag each 'done' REAL, MOCK or UNPROVEN. Give the path or address of each file you saved."), Standards ("Name, by tag, the standards you followed and where your work shows each one. Name any standard you set aside and the person who allowed it. An agent cannot allow it."), Capture ("an address, a tool on the user's side, a preference, a trap. Put it in the project folder with where it came from and its date. A gap or conflict in a standard is a finding. What a step taught goes in its Learned line. Never write a secret. List what you wrote under Captures, last in your reply."); who runs it: Agent.Coder.RulesExecuted, Agent.Concept.RulesExecuted, Agent.Tester.RulesExecuted (When you hand back, run Kai-Zen.Return.); Kai-Zen step Report ("Take the work back and delegate the next work at once. Report to the requester when the run ends"; hands off to Kai-Zen.EvidenceRules) and Kai-Zen.StopRules.Owed (say what is done, what is not, and why); Jackfruit.Report; GilbPlatform.SignOff.Name ("End the message with the name of the standard you loaded, alone on the last line ... Do not sign a name you did not load. If you loaded no standard, the last line is [No agent — base assistant]."); Kai-Zen.StopRules.Finding; Agent.Coder.AfterIncrement (self-QA, screenshot path and count in your return), Agent.Tester.Task (Return the verdict and the report's path or address. Always both.). GilbApp af0f04ee838a4d55887d2ff1f72edd49a4ff7ca0 (version.json read 2026-11-10), code read: components/agent/AgentChat.tsx:850-880 (a reply is drawn as plain text, `whitespace-pre-wrap`, no links; under it `{model || "Claude Code"} · your Claude subscription` or `{model || "SuperGrok"} · your SuperGrok` only when your own AI paid), :67 and :603-604 (MAX_TURNS = 32, the thread is cut to the last 32 turns), :937-944 with lib/base/agentChat.ts:120-143 (the standing notice above the message box); AgentChat.tsx:4-5 (thread in this tab). Kai-Zen (10 steps) and Jackfruit (13 steps) have no such step (Jackfruit.Report lists vaults, accesses and records); Kai-Zen.Sign and Jackfruit.Sign both hand off to GilbPlatform.SignOff; Kai-Zen.Return (Report, Standards, Capture). Also: Kai-Zen.ValueDeliveryStepRules.Learning, Agent.Coder, Agent.Concept, Agent.Tester.
Checked against GilbApp d529620.