WEBVTT

1
00:00:00.000 --> 00:00:02.300
An agent is a model using tools in

2
00:00:02.350 --> 00:00:02.550
a loop

3
00:00:02.860 --> 00:00:05.810
to achieve complex tasks that require multiple

4
00:00:05.860 --> 00:00:06.595
steps.

5
00:00:06.670 --> 00:00:09.630
Dominik Lukeš closed his Oxford Mapping the

6
00:00:09.680 --> 00:00:12.560
Landscape series with this definition, which is not

7
00:00:12.610 --> 00:00:14.760
official but one most people in the field

8
00:00:14.810 --> 00:00:15.910
would accept.

9
00:00:16.130 --> 00:00:18.776
The keywords are tools and loop.

10
00:00:18.850 --> 00:00:20.930
A chatbot can tell you how to fill

11
00:00:20.980 --> 00:00:22.850
in an expenses form.

12
00:00:23.030 --> 00:00:25.440
An agent does the steps.

13
00:00:25.670 --> 00:00:27.960
When the University of York sent Dominik a

14
00:00:28.010 --> 00:00:30.380
claim form, he asked his agent to find

15
00:00:30.430 --> 00:00:32.043
the receipts in his emails,

16
00:00:32.120 --> 00:00:34.389
compile the form and attach copies.

17
00:00:34.460 --> 00:00:36.830
It also filled in a Word form, and

18
00:00:36.880 --> 00:00:38.747
it took about thirty minutes.

19
00:00:38.820 --> 00:00:41.340
The word agent usually suggests a secret agent

20
00:00:41.390 --> 00:00:44.340
or a customer service agent, but Dominik finds

21
00:00:44.390 --> 00:00:46.663
a real estate agent a better comparison.

22
00:00:46.740 --> 00:00:48.520
Such an agent knows the rules and the

23
00:00:48.570 --> 00:00:50.790
local market, understands what people want from a

24
00:00:50.840 --> 00:00:53.290
home, follows instructions and uses

25
00:00:53.340 --> 00:00:55.890
databases, Excel and email.

26
00:00:56.050 --> 00:00:58.560
A language model already has the knowledge, the

27
00:00:58.610 --> 00:01:00.517
understanding and the communication.

28
00:01:00.590 --> 00:01:03.180
What it lacks is the ability to use

29
00:01:03.230 --> 00:01:03.864
tools.

30
00:01:03.940 --> 00:01:06.450
Every AI app is a model plus

31
00:01:06.500 --> 00:01:09.790
orchestration, the code around the model.

32
00:01:10.240 --> 00:01:12.690
The history of agents is the model taking

33
00:01:12.730 --> 00:01:14.060
over more of it.

34
00:01:14.570 --> 00:01:17.740
In November 2022 a chatbot only

35
00:01:17.790 --> 00:01:18.860
answered.

36
00:01:18.970 --> 00:01:21.602
Then it gained search and file tools.

37
00:01:21.680 --> 00:01:24.050
Then you could delegate a whole task, which

38
00:01:24.100 --> 00:01:25.720
took much longer.

39
00:01:26.150 --> 00:01:28.199
Then tasks ran for hours,

40
00:01:28.270 --> 00:01:30.505
and now some agents are always on.

41
00:01:30.580 --> 00:01:33.680
Until 2023 an agent meant a program following

42
00:01:33.730 --> 00:01:35.636
rules written by hand.

43
00:01:35.710 --> 00:01:39.030
The 2010 textbook by Russell and Norvig describes

44
00:01:39.080 --> 00:01:42.260
such agents, but Dominik says that research mostly

45
00:01:42.310 --> 00:01:44.730
did not carry over to language models.

46
00:01:45.240 --> 00:01:48.490
Power Automate flows work this way: when an

47
00:01:48.540 --> 00:01:51.028
email arrives, file the attachment.

48
00:01:51.100 --> 00:01:53.180
A model could be added as one step,

49
00:01:53.230 --> 00:01:55.960
to classify or draft, but what happened next

50
00:01:56.010 --> 00:01:57.296
was still written by a person.

51
00:01:57.370 --> 00:02:00.081
Self-orchestration began with tool calling.

52
00:02:00.160 --> 00:02:03.480
In 2023 ChatGPT plugins let the model

53
00:02:03.530 --> 00:02:05.020
choose when to use one of three that

54
00:02:05.070 --> 00:02:06.385
were switched on,

55
00:02:06.460 --> 00:02:08.830
and models learned to write code for a

56
00:02:08.880 --> 00:02:10.489
sum instead of guessing.

57
00:02:10.560 --> 00:02:11.955
It was unreliable.

58
00:02:12.030 --> 00:02:14.550
Google said in June 2023 that Bard was

59
00:02:14.600 --> 00:02:17.260
thirty per cent more accurate on maths problems,

60
00:02:17.310 --> 00:02:19.111
which still left many wrong answers.

61
00:02:19.190 --> 00:02:20.900
The loop came early too.

62
00:02:21.020 --> 00:02:24.190
BabyAGI in March 2023 had

63
00:02:24.240 --> 00:02:27.200
three agents, each a single model call, inside

64
00:02:27.250 --> 00:02:29.446
about a hundred and forty lines of Python.

65
00:02:29.520 --> 00:02:32.180
Everyone was excited, but it was

66
00:02:32.230 --> 00:02:34.530
brittle, because the models were not reliable enough

67
00:02:34.580 --> 00:02:37.060
to behave well across long sequences.

68
00:02:37.290 --> 00:02:39.603
The Python loop was the intelligence.

69
00:02:39.680 --> 00:02:42.210
Reasoning models arrived in late 2024,

70
00:02:42.760 --> 00:02:45.270
and in 2025 a model could call a

71
00:02:45.320 --> 00:02:47.519
search in the middle of its thinking.

72
00:02:47.590 --> 00:02:50.620
That took off in coding, where an agent

73
00:02:50.670 --> 00:02:53.480
plans files, writes them, runs them and fixes

74
00:02:53.530 --> 00:02:54.189
errors.

75
00:02:54.260 --> 00:02:56.270
Code suits a loop because it can be

76
00:02:56.320 --> 00:02:57.200
tested.

77
00:02:57.420 --> 00:03:00.220
By mid-2025 people were using the same tools

78
00:03:00.280 --> 00:03:01.920
to edit notes and emails.

79
00:03:02.000 --> 00:03:05.170
Then in late November 2025 Claude Opus

80
00:03:05.220 --> 00:03:08.250
4.5, GPT-5.2 and Gemini 3

81
00:03:08.300 --> 00:03:10.862
Flash made long loops reliable.

82
00:03:10.940 --> 00:03:13.450
They were trained for planning and tool use,

83
00:03:13.500 --> 00:03:16.460
through reinforcement learning on real computers, so

84
00:03:16.510 --> 00:03:19.760
orchestration once written as scripts is moving

85
00:03:19.810 --> 00:03:20.684
into training.

86
00:03:20.760 --> 00:03:23.420
New versions now arrive every six to eight

87
00:03:23.470 --> 00:03:25.009
weeks, and the task length

88
00:03:25.080 --> 00:03:27.160
a model can finish half the time reached

89
00:03:27.210 --> 00:03:29.151
about seventeen hours of human work.

90
00:03:29.230 --> 00:03:31.423
The loop itself has not changed.

91
00:03:31.500 --> 00:03:34.230
The harness, the app on your computer, sends

92
00:03:34.280 --> 00:03:35.674
instructions to the model.

93
00:03:35.750 --> 00:03:38.480
The model plans and sends commands back, the

94
00:03:38.530 --> 00:03:40.610
harness passes them to the computer, and the

95
00:03:40.660 --> 00:03:42.784
results return to the model.

96
00:03:42.860 --> 00:03:45.370
Dominik compares the model to a brain, the

97
00:03:45.420 --> 00:03:47.650
harness to hands and the tools to a

98
00:03:47.700 --> 00:03:49.110
notebook and pen.

99
00:03:49.460 --> 00:03:52.421
As models improve, the harness shrinks.

100
00:03:52.500 --> 00:03:54.400
An example shows the power of this.

101
00:03:54.470 --> 00:03:57.360
Asked to find last year's evaluation numbers, remembered

102
00:03:57.410 --> 00:03:59.850
as one person's work from February, a person

103
00:03:59.900 --> 00:04:01.730
opens file after file.

104
00:04:01.810 --> 00:04:05.130
The agent treats the remembered details as clues,

105
00:04:05.180 --> 00:04:07.400
searches many things at once, compares what it

106
00:04:07.450 --> 00:04:07.887
reads,

107
00:04:07.960 --> 00:04:10.840
and returns 148 responses with an average rating

108
00:04:10.890 --> 00:04:13.850
of four point three out of five, with

109
00:04:13.900 --> 00:04:15.170
its sources.

110
00:04:15.880 --> 00:04:18.073
It had corrected the person and the month.

111
00:04:18.150 --> 00:04:20.300
A model writes from its training and its

112
00:04:20.350 --> 00:04:23.670
context: the prompt, the chat history and any

113
00:04:23.720 --> 00:04:25.770
files, attached or found.

114
00:04:25.850 --> 00:04:28.140
There is nothing else, so the model itself

115
00:04:28.190 --> 00:04:29.830
never remembers you.

116
00:04:29.950 --> 00:04:33.090
Memory is notes saved in files.

117
00:04:33.390 --> 00:04:35.400
That is why a desktop agent works on

118
00:04:35.450 --> 00:04:37.960
files in a folder, with standing instructions in

119
00:04:38.010 --> 00:04:39.960
a file called AGENTS.md,

120
00:04:40.870 --> 00:04:43.310
and skills, which are instructions it reads only

121
00:04:43.360 --> 00:04:44.712
when a task matches.

122
00:04:44.790 --> 00:04:48.157
Connectors give agents tools for email and calendars.

123
00:04:48.230 --> 00:04:51.190
Context can live in folders, GitHub, databases or

124
00:04:51.240 --> 00:04:53.680
Notion, and Dominik advises keeping it somewhere that

125
00:04:53.730 --> 00:04:56.540
belongs to you rather than inside one chatbot.

126
00:04:57.390 --> 00:05:00.060
Agents can also work in teams: a lead

127
00:05:00.110 --> 00:05:03.130
agent starts others, as Codex and Claude Code

128
00:05:03.190 --> 00:05:03.866
often do.

129
00:05:03.940 --> 00:05:06.290
Choosing a model matters more for agents.

130
00:05:06.430 --> 00:05:09.680
It needs good judgement, tool choice, planning,

131
00:05:09.730 --> 00:05:12.470
coherence over long loops and management of its

132
00:05:12.520 --> 00:05:13.790
own context,

133
00:05:14.500 --> 00:05:16.280
which is not the list for the best

134
00:05:16.330 --> 00:05:17.750
model overall.

135
00:05:17.940 --> 00:05:21.120
On long office tasks Anthropic's models take the

136
00:05:21.170 --> 00:05:24.310
first three places in Artificial Analysis's tests,

137
00:05:24.760 --> 00:05:28.360
while Google's best usable model is fifty-third.

138
00:05:28.720 --> 00:05:31.160
Dominik would not trust a local open model

139
00:05:31.210 --> 00:05:32.452
with anything important,

140
00:05:32.530 --> 00:05:34.970
because in a long loop every later step

141
00:05:35.020 --> 00:05:36.662
builds on an early mistake.

142
00:05:36.740 --> 00:05:40.130
The agents people hear about, deep research, coding,

143
00:05:40.180 --> 00:05:42.990
computer and browser use, workflows and personal

144
00:05:43.040 --> 00:05:46.220
assistants, all run this loop and have merged

145
00:05:46.270 --> 00:05:47.170
into one.

146
00:05:47.370 --> 00:05:50.170
They run on your computer, inside apps, or

147
00:05:50.220 --> 00:05:53.330
on their own cloud computer, and Dominik

148
00:05:53.380 --> 00:05:56.700
recommends starting on your computer, where you have

149
00:05:56.750 --> 00:05:57.514
most control.

150
00:05:57.590 --> 00:05:59.220
The slides he had no time for list

151
00:05:59.270 --> 00:06:00.840
what can go wrong when an agent works

152
00:06:00.890 --> 00:06:02.190
on your files:

153
00:06:02.570 --> 00:06:05.750
unintended edits or deletions, and

154
00:06:05.800 --> 00:06:07.530
changes made in bulk.

155
00:06:07.780 --> 00:06:10.150
The advice is to keep backups, ask for

156
00:06:10.200 --> 00:06:12.710
a changelog, and read each permission request before

157
00:06:12.760 --> 00:06:13.304
answering.

158
00:06:13.380 --> 00:06:16.190
For the future, agents in science now propose

159
00:06:16.240 --> 00:06:19.490
hypotheses and experiments while people still

160
00:06:19.540 --> 00:06:20.707
run and judge them.

161
00:06:20.780 --> 00:06:23.960
Open questions remain: whose agent this is, how

162
00:06:24.010 --> 00:06:26.670
an agent pays, how agents from different companies

163
00:06:26.720 --> 00:06:27.377
work together,

164
00:06:27.450 --> 00:06:29.600
who answers for what an agent did, and

165
00:06:29.650 --> 00:06:30.602
how to measure them.

166
00:06:30.680 --> 00:06:32.460
Dominik warns that this may be out of

167
00:06:32.510 --> 00:06:34.950
date in six months, but the loop will

168
00:06:35.000 --> 00:06:35.550
be with us.

169
00:06:35.630 --> 00:06:38.290
To get ready, he suggests collecting more material,

170
00:06:38.340 --> 00:06:39.948
writing down institutional knowledge,

171
00:06:40.020 --> 00:06:42.020
and putting as few barriers as possible between

172
00:06:42.070 --> 00:06:44.300
agents and your data, in a way you

173
00:06:44.350 --> 00:06:47.160
can trust, because agents will succeed by getting

174
00:06:47.210 --> 00:06:48.190
context.

175
00:06:48.310 --> 00:06:50.820
Asked whether agents can keep improving, he pointed

176
00:06:50.870 --> 00:06:54.120
to long-running coherence, tool choice and context

177
00:06:54.170 --> 00:06:54.901
management,

178
00:06:54.980 --> 00:06:57.200
which is why people noticed over a month

179
00:06:57.250 --> 00:06:58.991
that they could trust their agents more.